HAL Id: tel-01228850 https://pastel.archives-ouvertes.fr/tel-01228850 Submitted on 13 Nov 2015 HAL is a multi-disciplinary open access archive for the deposit and dissemination of sci- entific research documents, whether they are pub- lished or not. The documents may come from teaching and research institutions in France or abroad, or from public or private research centers. L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des établissements d’enseignement et de recherche français ou étrangers, des laboratoires publics ou privés. Dynamic resource allocation for cellular networks with interference Tania Villa Trapala To cite this version: Tania Villa Trapala. Dynamic resource allocation for cellular networks with interference. Networking and Internet Architecture [cs.NI]. Télécom ParisTech, 2013. English. NNT: 2013ENST0054. tel- 01228850
Transcript
HAL Id: tel-01228850https://pastel.archives-ouvertes.fr/tel-01228850
Submitted on 13 Nov 2015
HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.
L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.
Dynamic resource allocation for cellular networks withinterference
Tania Villa Trapala
To cite this version:Tania Villa Trapala. Dynamic resource allocation for cellular networks with interference. Networkingand Internet Architecture [cs.NI]. Télécom ParisTech, 2013. English. �NNT : 2013ENST0054�. �tel-01228850�
M. Tommy SVENSSON, Prof., Chalmers University of Technology Rapporteur
M. David GESBERT, Prof., EURECOM Examinateur
M. Ruben MERZ, Dr., Swisscom Examinateur
M. Antonio CIPRIANO, Dr., Thales Communications & Security Examinateur
TELECOM ParisTechécole de l’Institut Télécom - membre de ParisTech
DISSERTATIONIn Partial Fulfillment of the Requirements
for the Degree of Doctor of Philosophy
from TELECOM ParisTech
Specialization: Wireless Communications
Tania Villa Trapala
Dynamic resource allocation for cellular networks
with interference
Defense scheduled on the 26th of September 2013 before a committeecomposed of:
Reviewers Prof. Angel Lozano, Universitat Pompeu FabraProf. Tommy Svensson, Chalmers University of Technology
Examiner Prof. David Gesbert, EURECOMDr. Ruben Merz, SwisscomDr. Antonio Cipriano, Thales Communications & Security
Thesis Supervisor Prof. Raymond Knopp, EURECOM
THESEprésentée pour obtenir le grade de
Docteur de TELECOM ParisTech
Spécialité: Wireless Communications
Tania Villa Trapala
Gestion dynamique de ressources appliquée aux
réseaux cellulaires avec interférence
Soutenance de thèse prévue le 26 Septembre 2013 devant le jurycomposé de :
Rapporteurs Prof. Angel Lozano, Universitat Pompeu FabraProf. Tommy Svensson, Chalmers University of Technology
Examinateur Prof. David Gesbert, EURECOMDr. Ruben Merz, SwisscomDr. Antonio Cipriano, Thales Communications & Security
Directeur de thèse Prof. Raymond Knopp, EURECOM
Abstract
Mobile networks have experienced dramatic growth during the past decades.Modern communication systems require high data rates and better qualityof service control for services such as voice telephony, online gaming, webbrowsing, etc. The main obstacle found in wireless communication networksis the time-varying nature of the physical channel. Taking this into account,the goal of a system designer is to simplify the overall network design andoptimize the performance.
The aim of this thesis is to design, implement and evaluate practical cross-layer algorithms to handle interference and allocate the radio resources inan efficient way for LTE and post-LTE uncoordinated networks. We developmathematical and computational interference models that allow us to under-stand the behavior of such networks and we apply an information-theoreticapproach to different interference scenarios and traffic characteristics. Wehave tried to remain as close as possible to practical systems to be able totest the feasibility of the proposed techniques.
As a part of the evolution to 4G systems, the introduction of small-cellsthat overlay the existing cellular network has been envisioned to fill in thecoverage white spots or serve mobile and outdoor users where the cellularnetwork is not deployed.
This thesis deals with performance evaluation of interference scenariosin 4G networks, in particular those arising from small-cell deployments.The work in this thesis also deals with analysis of resource-allocation andincremental-redundancy based hybrid automatic repeat request (HARQ)with bursty interference (or more general time-varying channels) which al-lows for only partial channel state information at the transmitter. The workis then applied to practical scheduler design for LTE base stations and in-cludes performance analysis for real LTE modems.
We showed that, in general, by adapting the number of physical dimen-sions across rounds, we can exploit the interference mitigation effects ofHARQ using it not only to recover from errors but for interference cancel-lation. We proposed efficient resource allocation algorithms to increase thethroughput, which can potentially come very close to optimal performance.
3.1 Retransmission Markov chain Xn: a frame transmission at-tempt always initiates and finishes in state 0. A frame re-transmission corresponds to a transition from state i to i+1.A successful frame transmission corresponds to a transitionfrom any state i = 0, . . . ,Mmax to the state 0. Finally, theframe is dropped if state Mmax + 1 is reached. Note thatq[m] = 1− p[m], ∀m. . . . . . . . . . . . . . . . . . . . . . . 18
3.2 In (a) we have the CDF of the mutual information underRayleigh fading and without any interference, (b) shows theCDF of the mutual information with one interferer, and (c)for two interferers. The interferers have the same power asthe user of interest. . . . . . . . . . . . . . . . . . . . . . . . . 21
3.3 We consider different modulations, for HARQ and ARQ, withMmax = 3 retransmissions. We show the average throughputfor both retransmission protocols and we see the improvementin throughput of HARQ over ARQ for all modulations. . . . . 22
3.4 We consider QPSK modulation, for HARQ, and we show theaverage throughput for different numbers of retransmissionsMmax = 1, 2, 3, 5, 10. In (a) we plot the case without anyinterference and in (b) we see a different trend for the casewith two strong interferers, (b) shows that adapting the rateis an optimal way to increase the throughput. . . . . . . . . . 23
3.5 Manhattan-like topology with the user of interest at the edgeof the apartment and the rest of the interfering femtocellsplaced at the middle of the surrounding apartments . . . . . . 24
vii
viii List of Figures
3.6 We look at the scenario in figure 3.5 and we consider QPSKmodulation, R = 2, Mmax = 3 for HARQ and ARQ. Weshow the average throughput with no interference and withinterference canceling two interferers for an SNR from −10 to25 dB. We do a Gaussian approximation for the interferersthat are not canceled. The SNR without interference cancel-ing is −7 dB and 6 dB for interference cancellation of the twostrongest interferes. The corresponding throughputs are 0.15and 1, therefore there is a gain of ten times in throughput. . . 24
3.7 We consider QPSK modulation, we show the average through-put when the interference is randomized with an activity fac-tor µ = 0.5, 0.75, which means that the interferers will beactive either half or 75% of the time and we compare withthe case of the interference present all the time. We see thatactivity factor of 0.5 has the lowest throughput and that ac-tivity factor of 75% is closer to the corresponding curve forµ = 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.1 Figure (a) shows the interference scenarios for HetNets in theDL, figure (b) shows the interference scenarios for HetNets inthe UL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2 Sparse traffic in a delay-constrained scenario. Traffic arrivalsin the eNB MAC layer are sparse as depicted in blue (there arethree of them). The latency constraint is four slots, i.e. thereare up to four possible PDSCH channel allocations. Becauseof the sparse traffic, CQI is outdated or unavailable on thefirst slot. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.3 For different values of the probability of outage after the sec-ond round Pout,2, we calculate the corresponding SNR for thedifferent scenarios. The symbol λ is the correlation coefficientbetween the actual channel and the channel corresponding tooutdated CQI information. We compare a correlation coeffi-cient value of 50%, 10% and uncorrelated case. For compar-ison purposes, we also plot two more cases. First when noACK/NACK feedback is available from the HARQ process.Second, when Pout,1 is fixed to 50% with ρ = 0.5 to make surethat 50% of the dimensions are used in each round. . . . . . . 38
List of Figures ix
4.4 The axis on the left (solid lines) shows the spectral efficiencyversus SNR for the different scenarios. We set Pout,2 to 1%.The symbol λ is the correlation coefficient between the ac-tual channel and the outdated/stale CQI information. Wecompare correlation coefficients of 50%, 10% and uncorrelatedcase. For comparison purposes we plot the curve for the er-godic capacity and Pout,1 fixed to 50% with ρ = 0.5 to makesure that 50% of the dimensions are used in each round. Theaxis on the right (dashed lines) shows the ratio of dimensionsused in the two rounds. . . . . . . . . . . . . . . . . . . . . . 40
4.5 Coding Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 414.6 Downlink of a macrocell with a femtocell interfering. . . . . . 434.7 The axis on the right (solid lines) shows the zero-outage through-
put for the HARQ protocol with different number of rounds,while the axis on the left (dashed lines) shows the ratio of di-mensions per round for the three rounds, zero-outage HARQprotocol. In both cases the channel is AWGN with Gaussiansignals and there is one interferer with probability p = 0.5.The interference strength is the same as the user of interest(SNR1 = SNR2). . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.8 Throughput of the two rounds HARQ protocol in an AWGNchannel with Gaussian signals. There is one interferer withprobability p = 0.05, 0.5. . . . . . . . . . . . . . . . . . . . . 47
4.9 In (a) we have the reliable throughput for the HARQ proto-col with different number of rounds under Rayleigh fading,constant over rounds, and without interference, (b) shows thereliable throughput different number of rounds and one dom-inant interferer. The rates across rounds are fixed and theoperating SNR is 10 dB. . . . . . . . . . . . . . . . . . . . . . 51
4.10 In (a) we show the rate optimization of the HARQ protocol fora diferent number of rounds Mmax = 1, 2 in a Rayleigh fadingchannel with QPSK modulation. The rates are fixed acrossrounds R1 = R2 = R. Figure (b) shows the correspondingprobability of outage. . . . . . . . . . . . . . . . . . . . . . . . 53
4.11 In (a) we show the rate optimization of the HARQ protocolfor a diferent number of rounds Mmax = 1, 2 in a Rayleighfading channel with QPSK modulation. The rates are fixedacross rounds R1 = R2 = R. There is one interferer all thetime. Figure (b) shows the corresponding probability of outage. 54
4.12 In (a) we show the rate optimization with an outage constraintof 10%. The channel is constant across the HARQ rounds andthere is no interference. This is equivalent to a Non-Line-Of-Sight (NLOS) with slow fading channel. Figure (b) shows thecorresponding curves for an outage constraint of 1% . . . . . 56
x List of Figures
4.13 In (a) we show the rate optimization with an outage constraintof 10%. The channel is iid across the HARQ rounds and thereis no interference. In this case, it is equivalent to a NLOSchannel with fast fading or frequency hopping. Figure (b)shows the corresponding curves for an outage constraint of 1% 58
4.14 Rate optimization for Rayleigh fading on the downlink chan-nel. There is one dominant interferer with an activity factorof 50%. The outage constraint is 1% . . . . . . . . . . . . . . 59
4.15 Rate optimization for Rayleigh fading on the downlink chan-nel. There is one dominant interferer with an activity factorof 50%. The outage constraint is 10% . . . . . . . . . . . . . 59
4.16 Rate optimization for Rayleigh fading on the uplink channel.There is one dominant iid interferer. The outage constraint is10% . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
4.17 Manhattan-like topology with the user of interest at the edgeof the apartment and the rest of the interfering small-cellsplaced at the middle of the surrounding apartments . . . . . . 61
4.18 Downlink of a small-cell user. Interference is coming from theneighboring small-cells with an activity factor µ = 50%. Themodulation is QPSK, 16-QAM, or 64-QAM and the SNR isequal to the interference strength and is 10dB. We consider thestrongest interferer and we make a Gaussian approximationfor the rest. The rate is adapted across rounds for all cases. . 62
4.19 Downlink of a macrocell with a femtocell interfering. . . . . . 63
4.20 Downlink of a QPSK macro cell user under Rayleigh fading.Interference is coming from a 16-QAM femtocell active onlypart of the time. Curves are shown for different activity fac-tors 30%, 50% and 70%. . . . . . . . . . . . . . . . . . . . . . 63
5.1 In (a), we consider the scenario without CQI (uncorrelatedchannels), and we plot the rate in the first round (R1) fordifferent values of SNR. We fix the probability of outage afterthe second round to 1%. In (b), for the scenario withoutCQI (uncorrelated channels), we plot the ρ parameter againstdifferent values of SNR. We fix the probability of outage afterthe second round to 1%. ρ determines the rate used in thesecond round according to equation [4.6] . . . . . . . . . . . . 73
5.2 Probability of Outage after two HARQ rounds. The channelis AWGN and there is no interference. . . . . . . . . . . . . . 75
5.3 Probability of Outage after two HARQ rounds. There is oneinterferer of the same strength as the user of interest with a50% probability of being active. The channel is AWGN forboth users. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
List of Figures xi
5.4 Probability of Outage after the first round of the HARQ pro-tocol. The channel is AWGN and there is no interference. . . 78
5.10 Probability of Outage after the first round of the HARQ pro-tocol. There is one interferer of the same strength as the userof interest with a 50% probability of being active. The channelis AWGN for both users. . . . . . . . . . . . . . . . . . . . . . 85
5.11 Spectral efficiency for different resource allocations. There isone interferer of the same strength as the user of interest witha 50% probability of being active. The channel is AWGN forboth users. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5.13 Probability of Outage after the first round of the HARQ pro-tocol. There is one interferer of the same strength as the userof interest with a 50% probability of being active. The channelis AWGN for the user of interest and Rayleigh for the interferer. 88
5.14 Spectral efficiency for different resource allocations. There isone interferer of the same strength as the user of interest witha 50% probability of being active. The channel is AWGN forthe user of interest and Rayleigh for the interferer. . . . . . . 89
5.15 Probability of Outage after the first round of the HARQ pro-tocol. There is one interferer of the same strength as the userof interest with a 50% probability of being active. Both theuser of interest and the interferer experience a Rayleigh fadingchannel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.16 Spectral efficiency for different resource allocations. There isone interferer of the same strength as the user of interest witha 50% probability of being active. Both the user of interestand the interferer experience a Rayleigh fading channel. . . . 91
5.17 Probability of Outage after the first round of the HARQ pro-tocol. There is one interferer of the same strength as the userof interest with a 50% probability of being active. The chan-nel model is EPA for the user of interest and AWGN for theinterferer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
xii List of Figures
5.18 Spectral efficiency for different resource allocations. There isone interferer of the same strength as the user of interest witha 50% probability of being active. The channel model is EPAfor the user of interest and Rayleigh for the interferer. . . . . 93
dans le DL, la figure (b) montre les scénarios d’interférencepour HetNets dans le UL . . . . . . . . . . . . . . . . . . . . . 118
A.8 Le trafic épars dans un scénario de retard limité. Les ar-rivées du traffic dans la couche MAC eNB sont rares commereprésenté en bleu (il y a trois d’entre eux). La contraintede latence est de quatre emplacements, à savoir il ăăăă sontjusqu’à quatre attributions de canaux PDSCH possibles. Enraison de la circulation clairsemée, le CQI n’est pas à jour oun’est pas disponible sur la première tranche. . . . . . . . . . . 119
Here are the main acronyms used in this document. The meaning of anacronym is usually indicated once, when it first appears in the text.
3G Third Generation4G Fourth Generation3GPP Third Generation Partnership ProjectABSF Almost Blank SubframeAMC Adaptive Modulation and CodingAWGN Additive White Gaussian NoiseBLER Block Error RateCDF Cumulative Distribution FunctionCP Cyclic PrefixCQI Channel Quality IndicatorCSI Channel State InformationDCI Downlink Control InformationDL DownlinkDRX Discontinuous ReceptionDSL Digital Subscriber LineEPA Extended PedestrianeNB eNodeB (LTE base station)HeNB Home-eNodeB (or femtocell)HetNet Heterogeneous NetworkHARQ Hybrid Automatic Repeat Requestiid Independent and Identically DistributedIoT Internet of ThingsIR Incremental RedundancyLTE Long Term EvolutionM2M Machine-to-MachineMAC Multiple Access ChannelMCS Modulation and Coding SchemeMIESM Mutual Information Effective SINR MappingMIMO Multiple Input Multiple OutputNDI New Data Indicator
xiii
xiv Acronyms
NLOS Non-Line of SightOAI OpenAirInterfaceOFDMA Orthogonal Frequency Division Multiple AccessPDU Packet Data UnitPDCCH Physical Downlink Control ChannelPDSCH Physical Downlink Shared ChannelPRB Physical Resource BlockPUCCH Physical Uplink Control ChannelPUSCH Physical Uplink Shared ChannelQoS Quality of ServiceQPSK Quadrature Phase-Shift KeyingRB Resource BlockRBG Resource Block GroupRE Resource ElementRLC Radio Link ControlSC-FDMA Single Carrier Frequency Division Multiple AccessSDR Software-Defined RadioSINR Signal to Interference and Noise RatioSISO Single Input Single OutputSNR Signal to Noise RatioSRS Sound Reference SignalTBS Transport Block SizeTTI Transmission Time IntervalUE User EquipmentUL Uplink
Notations
Here is a list of the main operators and symbols used in this document. Wehave tried to keep the notation consistent throughout the thesis, but rarelysymbols have different definitions in different chapters and in that case theyare defined very explicitly to avoid any confusion.
x or X Input signalh,H Channelz,Z Noiseσ2 Variance of the noisey,Y Received signal|x| Absolute value of scalarE Palm ExpectationE [.] Mathematical expectationMmax Maximum number of HARQ roundsCN Complex Normal DistributionH (.) = −E log p (.) Entropy of the argumentFχ (χ) Cumulative distribution function of χQM (a, b) Marcum Q-function with M degrees of freedom and paramters a
and blog2 All logarithms are to the base 2I(.; .) Mutual information between the two argumentsµ Activity factorR̄ Spectral effiency or throughputρ Ratio of physical dimensions used across HARQ roundsℑ{} Imaginary partΦI(ω) Characteristic function of the mutual information Iβ1,2 Calibration factorsS1 |x1| i.e. cardinality of the constellation of x1S2 |x2| i.e. cardinality of the constellation of x2
xv
xvi Notations
Chapter 1
Introduction
Modern communication systems, like the Long Term Evolution (LTE) stan-dard [6], require high data rates and better Quality of Service (QoS) controlfor services such as voice telephony, online gaming, web browsing, etc. Inorder to cope with the requirements of these new types of services whilesimultaneously offering high data rates, the standards have to evolve andadapt. Both the challenges from the Physical layer (PHY) and the QoS de-mands from the applications have to be taken into account. Future wirelesscommunications networks require optimization of parameters in all layers.In a cross-layer design, rate, power and coding at the PHY can be adaptedto meet the requirements of the applications given the current channel andnetwork conditions (See figure 1.1).
By designing policies that combine interference mitigation at the PHYlayer with scheduling algorithms and rate adaptation at the Medium AccessControl layer (MAC), and further radio resource management at the RadioLink Control layer (RLC), we can obtain higher throughput, more bandwidthand therefore, more efficient networks.
As a part of the evolution to Fourth Generation (4G) systems, the in-troduction of small-cells that overlay the existing cellular network has beenenvisioned to fill in the coverage white spots or serve mobile and outdoorusers where the cellular network is not deployed. However, since these small-cells may not be connected to the operator backhaul network, coordinationbetween them for resource management is hardly feasible (see figure 1.2).
The aim of this thesis is to design, implement and evaluate practicalcross-layer algorithms to handle interference and allocate the radio resourcesin an efficient way for LTE and post-LTE uncoordinated networks. We de-velop mathematical and computational interference models that allow us to
1
2 Chapter 1 Introduction
Figure 1.1: Cross-Layer Techniques.
understand the behavior of such networks. We first do a theoretical study byapplying an information-theoretic approach to different interference scenar-ios and traffic characteristics. We have tried to remain as close as possible topractical systems to be able to test the feasibility of the proposed techniques.Secondly, we perform a full simulation study which ends in the implemen-tation and evaluation of the proposed techniques in the OpenAirInterfaceplatform (OAI) [3].
1.1 Contributions and Thesis Outline
The main obstacle found in wireless communication networks is the time-varying nature of the physical channel. Taking this into account, the goalof a system designer is to make the PHY/MAC smarter to simplify theoverall network design and optimize the performance. In our strategies, wehave tried to take into account scenarios and parameters that make themapplicable to practical systems.
In order to better understand the relevance of our study, we begin in
Figure 1.2: Femtocells are an example of small-cell deployments [4].
1.1 Contributions and Thesis Outline 3
Chapter 2 by describing the evolution of wireless communication networks,together with the new interference scenarios resulting from this evolution.We then explain the fundamentals of LTE and interference networks and wealso describe the basics of scheduling and link adaptation in LTE.
Chapter 3 deals with performance evaluation of interference scenariosin 4G networks, in particular those arising from femtocell deployments. Inthis chapter, we analyze the throughput of the network with the help ofan inhomogeneous discrete-time Markov chain and basic information theoryquantities. We consider one or two dominant interferers and we study a de-centralized interference mitigation scheme that combines Hybrid AutomaticRepeat and Request (HARQ) and Incremental Redundancy (IR) with aninterference cancellation decoder. For comparison purposes, we also studyan ARQ scheme where the information is not accumulated across transmis-sion rounds. Our performance evaluation based on analytical modeling andMonte Carlo evaluation of throughput shows that our scheme is effective atcombating interference without requiring any coordination. The results werepublished in
• Villa, Tania; Merz, Ruben; Knopp, Raymond, “Interference man-agement in femtocell networks with Hybrid-ARQ and inter-ference cancellation”, in the proceedings of IEEE Asilomar Con-ference on Signals, Systems, and Computers, November 2011, PacificGrove, CA, USA.
The fourth chapter starts by describing some of the key emerging applica-tions where our information-theoretic analysis can be applied. This chapteris crucial for understanding the importance of dynamic resource allocationschemes in LTE. Unlike any previous work, we consider the allocation ofphysical resources not fixed across HARQ transmissions. The latter is a realpossibility in schedulers for LTE base stations (eNodeB or eNB) and, to thebest of our knowledge, no well-known methodology exists for adapting phys-ical resources across HARQ rounds when subject to time-varying channelseither due to fading or time-varying interference or a combination of both.By doing this, we exploit the interference mitigation effects of HARQ usingit not only to recover from errors but for interference cancellation and wepropose efficient resource allocation algorithms to increase the throughput,which can potentially come very close to optimal performance.
We first present an analysis of interference-free networks with time-varying channels. Rather than performing extensive simulations, we take aninformation theoretic approach to derive analytical expressions that repre-sent the long-term throughput of a point-to-point link and consider practicalcases where there is a constraint on the outage probability representing thelatency of the protocol. We consider Gaussian input signals and we consider
4 Chapter 1 Introduction
cases where Channel Quality Indicator (CQI) feedback is either unavailable,or outdated.
The fourth chapter then examines the case of networks with interference.We motivate the use of inter-round resource allocation through a simple butillustrative analysis with Gaussian signals and interference. We include theuse of activity factors which model sporadic interference patterns charac-teristic of future heterogeneous networking deployments, in particular theinterference seen from small-cell base stations with bursty traffic in the re-ceiver of a macrocell user.
Finally, we look at practical interference scenarios to illustrate the appli-cations of our analytical framework. We model a Manhattan-like topologywhich represents a block of apartments with femtocells creating interference.We then model a macro-femto scenario where a macrocell is overlaid by afemtocell and we look at the downlink (DL) channel of the macrocell userwhen the interference is coming from the interfering femtocell. With the useof an activity factor we model the fact that the femtocell is not active allthe time. Finally, we explain the procedure that has to be followed to per-form PHY abstraction with the use of our analytical framework, given theimportance to accurately model the link performance in order to speed-upsimulations. The results were published in
• Villa, Tania; Merz, Ruben; Knopp, Raymond, “Adaptive modu-lation and coding with Hybrid-ARQ for latency-constrainednetworks”, in the proceedings of IEEE European Wireless Conference(EW2012), April 2012, Poznan, Poland.
• Villa, Tania; Merz, Ruben; Knopp, Raymond, “Adaptive transmis-sion and mutiple-access for sparse-traffic sources”, in the pro-ceedings of IEEE European Signal Processing Conference (EUSIPCO),August 2012, Bucharest Romania.
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation in heterogeneous networks”, in the proceedings of IEEEGlobal Communications Conference (GLOBECOM), December 2013,Atlanta, USA.
and has been accepted for publication in
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation for time-varying channels in next generation cellu-lar networks, Part I: a mathematical framework”, submitted toIEEE Transactions on wireless communications
1.1 Contributions and Thesis Outline 5
and will be submitted as one part of
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation for time-varying channels in next generation cellu-lar networks, Part II: applications in LTE”, under preparation.
Chapter 5 deals with practical scheduler design for LTE base stations.LTE offers a lot of flexibility in terms of resource allocation and, in partic-ular, resource allocation algorithms can be tailored for a particular class oftraffic with specific requirements. Nevertheless, work is still needed to ex-ploit this flexibility efficiently for key emerging applications. In this chapter,we study the performance of our dynamic resource allocation polices for thescheduling of IR-HARQ transmissions under the constraints of LTE coded-modulation. In this capther, we show the results of the implementation ofthe scheduler in the OAI software-defined radio (SDR) platform [3] in orderto test the performance and compliance of our resource allocation strategiesin LTE. We show that the implementation of such policies in a real system isfeasible withouth requiring any coordination or complex and time-consumingoptimization procedure. We show that our scheduling techniques work fordifferent environments and most importantly, we show that the results are inagreement with the theoretical results presented in Chapter 4. The resultswill be submitted as one part of
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation for time-varying channels in next generation cellu-lar networks, Part II: applications in LTE”, under preparation.
6 Chapter 1 Introduction
Chapter 2
Background
Mobile networks have experienced dramatic growth during the past decades.Third Generation systems (3G) are expanding the possibilities of informa-tion transfer and communication. Besides high data rate, 3G systems alsoenvisioned providing better QoS control for a variety of applications, fromvoice telephony and gaming, to web browsing, e-mail, and streaming multi-media applications. With the roll-out of 3G technology in many countries,researchers and standardization organizations have increased their efforts tooffer even higher data rates and more services to the users than 3G. Thecapability of high data rate transmission determines the kind of service thatcan be provided to users and the QoS they receive. Users now expect tohave the same on-demand access to multimedia content from anywhere andwhile “on the move”.
2.1 Evolution of Wireless Communication Systems
The evolution of wireless communication systems is mainly driven by the in-troduction of new services and the availability of more advanced technologies.Over the last couple of decades, cellular networks have grown exponentiallyand the demand for new and improved services has become an importantissue for operators. There is a need for new technologies to alleviate thecapacity limitations of the network and to maintain the QoS demanded bythe users. This has motivated the development of new standards like theThird Generation Partnership Project’s (3GPP) LTE standard [6] in orderto provide higher data rates and an improved QoS in wireless networks.
The 3GPP LTE standard has chosen Orthogonal Frequency DivisionMultiple Access (OFDMA) as the underlying modulation technology [7]. It is
7
8 Chapter 2 Background
a spectrally efficient version of multi-carrier modulation, where the subcarri-ers are selected in such a way that they are all orthogonal to one another overthe symbol duration. Physical layer DL transmissions are implemented us-ing OFDMA while Uplink (UL) transmission uses Single Carrier FrequencyDivision Multiple Access (SC-FDMA). The main difference between bothschemes is that in OFDMA data detection is performed in the frequency do-main while in SC-FDMA it is done in the time domain averaging the noiseover the entire bandwidth [52].
Despite being in the early days of roll-out, LTE has become the fastestdeveloping system due to the fast speeds and high quality user experiencethat it offers. It has now been launched on all continents, by 156 operatorsin 67 countries. By 2013, there are 6.4 billion of mobile subscriptions glob-ally [1]. Current commercial LTE deployments are based on 3GPP Release8 and 9 [5].
Data traffic will continue to grow, along with mobile data subscriptionsand an increase in the average data volume per subscription. In fact, overallmobile data traffic is expected to continue the trend of doubling each year [1].This growth in traffic and services will bring new technical challenges to theoperators, interference being one of the most performance-limiting.
The demand for new mobile services and for higher peak bit rates andsystem capacity is tackled with the evolution of the technology to 4G. The4G systems are based on 3GPP LTE and are progressing on a large scale,with 55 million users as of 2012 and from 1.6 to 2 billion users anticipatedin 2018 [1].
2.2 Interference Scenarios in 4G Networks
During the past 20 years, there has been a massive growth in traffic volume,number of devices connected, and an increased demand for video data. Fu-ture cellular networks should be able to cope with this increased demandand handle all the traffic in an efficient way.
There are new technical challenges, and potential interference scenariosthat vary with the type of deployments, requirements, high data rate andQoS levels. These new interference scenarios have been considered duringthe LTE Release 10 standardization:
• Macro-picocell interference (see figure 2.1).
• Macro-Home-eNodeB (HeNB) interference (see figure 2.2).
Enhancements in the operation of the base stations, advanced terminalreceivers and a new carrier type with reduced transmission of “always-on”signals are a requirement for high network energy efficiency [28]. With thetransmission of control signaling, a non-negligible amount of energy is usedby the power amplifier. Minimizing the transmission of “always-on” signals
2.2 Interference Scenarios in 4G Networks 9
Figure 2.1: Macro-eNB vs pico-eNB.
Figure 2.2: Macro-eNB vs HeNB.
allows the base station to turn off circuitry if it has no data to transmit. Byeliminating unnecessary transmissions, the interference is reduced, leading toimproved data rates in the network. Another aspect to consider is that withthe use of small-cells, sporadic traffic, and a reduction in the transmissionof control signals, interference becomes time-varying. When designing newtechniques, these new characteristics and requirements have to be taken intoaccount.
2.2.1 Heterogeneous Networks and Interference
As the traffic demand grows and the RF environment changes, the networkrelies on cell splitting or additional carriers to overcome capacity and linkbudget limitations and maintain uniform user experience. Moreover, siteacquisition for macro base stations with towers becomes more difficult indense urban areas. A more flexible deployment model is needed for operatorsto improve broadband user experience in a ubiquitous and cost-effective way.
The concept of Heterogeneous Networks (HetNets) has been introducedin LTE to address the capacity and coverage challenges resulting from theenormous and continuous growth of data services. The traditional macro net-work is deployed to provide umbrella coverage and smaller nodes are added asan underlay network to alleviate coverage holes and traffic hot zones. These
10 Chapter 2 Background
low-power nodes provide very high traffic capacity and user throughput lo-cally, for ex. indoor and outdoor hotspot positions. The macro layer ensuresservice availability and QoS over the entire coverage area. HetNets includemicrocells, picocells, femtocells, and distributed antenna systems (remote ra-dio heads), which are distinguished by their transmit powers, coverage areas,physical size, backhaul and propagation characteristics [33].
Energy-efficient load-balancing can be achieved by turning off the low-power nodes when there is no ongoing data transmission. A macro base-station can use Almost Blank Subframes (ABSF) to reserve some subframesfor small-cells [42].
2.2.2 Femtocells
A femtocell, or HeNB, is a low power, low cost wireless access point installedby the end user to improve voice and data performance. They are deployedon top of the macro network and its primary purpose is to provide enhancedcapacity for busy outdoor areas and improved coverage for indoor areas (seefigure 2.3). HeNBs connect mobile devices to the operator network throughthe broadband connection of the user. Since a femtocell can operate usingthe licensed spectrum of the operator, it can allow operators to guarantee anacceptable level of QoS to the user, both in terms of coverage and capacity,avoiding the need for additional network interfaces. Since HeNBs requireonly low transmission power, battery life of mobile phones increases and theinterference decreases.
The installation for such devices is planned to be plug and play, so theuser does not have to worry about cumbersome installation, and configura-tion. Updates and maintenance to the software and configuration will bedone transparently and controlled by operators.
2.2.3 Machine-to-Machine Communications
Over the last years, there has been a dramatic growth in communication be-tween devices. Their traffic can be characterized as small, delay-tolerant datapackets which are sent infrequently [61]. Applications of Machine-to-Machinecommunications (M2M) include those related to public safety, surveillancecameras, sensoring, monitoring, etc. M2M communications imply some tech-nical and non-technical challenges to the network such as:
• Allowing for very low-cost of device types
• Achieving low device energy consumption to ensure long battery lifefor relevant applications
• Providing extended coverage options in challenging locations
• Handling a very large number of devices per cell
2.3 Scheduling and Link Adaptation in LTE 11
Figure 2.3: Macrocell network overlaid by HeNBs.
With the introduction of a large amount of communicating devices through-out the network, interference is increased. As a mean to control inter-cellinterference, scheduling algorithms can be used together with link adaptationto adapt to a changing-channel and also to increase the throughput.
2.3 Scheduling and Link Adaptation in LTE
In LTE, the scheduling algorithms and rate adaptation at the MAC layer canbe combined with radio resource management at the RLC layer to obtainhigher throughput, more bandwidth and therefore, more efficient networks.Implementing efficient algorithms for radio resources management, packetscheduling, admission control or power and interference control are importantto optimize the capacity and performance.
Scheduling consists on allocating the transmission resources, Physical Re-source Blocks (PRBs) in LTE, to the users, every transmission opportunity.Given the variations experienced in the quality of a wireless channel, thechoice of other parameters such as Modulation and Coding Scheme (MCS)can be adapted with the goal of maximizing the capacity in the cell, whilesatisfying the QoS requirements of every user. In this way, the randomnessof the radio link can be taken into account and exploited to use the resourcesin the most efficient way. The scheduler interacts closely with the HARQmanager who is responsible for the scheduling of the retransmissions in case
12 Chapter 2 Background
of an incorrect reception. The LTE standard supports dynamic, channel-dependent scheduling to enhance overall system capacity.
In LTE, the scheduler resides at the eNB. Capacity is shared amongmultiple users on an on-demand basis. The purpose of the scheduler is todecide which terminal or base station transmit and on which set of resources.
Similar to OFDMA schedulers used on the DL, SC-FDMA schedulersfor the UL can be both time and frequency-opportunistic. An importantdifference between DL and UL scheduling is that CQI reporting is not neededsince the scheduler is located at the eNB which can measure UL channelquality through Sound Reference Signals (SRS) [61].
2.3.1 Resource Allocation in LTE
In LTE, the available bandwidth is divided into N subcarriers. From the Nsubcarriers, 12 or 24 adjacent subcarriers are grouped together forming whatis called a Resource Block (RB), which represents the minimal schedulingresource for both UL and DL transmissions and it corresponds to 180 KHz ofspectrum (see figure 2.4). LTE frames are divided into two slots of durationTslot = 0.5ms. A slot is formed by NRB RBs in the frequency domain for theduration of 6 or 7 OFDMA symbols in the time domain, depending on thelength of the Cyclic Prefix (CP) used. The CP is used for synchronizationpurposes and is attached to each slot. A specific subcarrier inside the RBis called a Resource Element (RE). Since the subcarriers in OFDMA areorthogonal there is no interference from within the cell, but interference isexperienced from the neighboring cells.
The number of RBs available depends on the bandwidth of the channel(see figure 2.5), and depending on the length of the CP, a different numberof OFDMA symbols is accommodated in a slot. Table 2.1 gives the differentnumber of RBs available for each of the channel bandwidths specified in the3GPP standard with the corresponding number of RBs.
Table 2.1: NRB vs Downlink System Bandwidth
Bandwidth 1.4 MHz 3 MHz 5 MHz 10 MHz 15 MHz 20MHz
NRB 6 15 25 50 75 100
Radio resource management aims at scheduling the available resources inthe best way to allow users to achieve a specific QoS. An intelligent mecha-nism has to consider the interference created with already assigned physicalresources.
2.3 Scheduling and Link Adaptation in LTE 13
Figure 2.4: LTE resource grid [9].
Figure 2.5: LTE bandwidth and resource blocks [28].
14 Chapter 2 Background
2.3.2 Discontinuous Reception (DRX)
Packet-data traffic is often highly bursty, with sporadic transmissions fol-lowed by inactivity periods. However, the control signals have to be moni-tored in order to receive UL grants or DL data transmissions and to adaptto the traffic variations. The latter consumes a non negligible amount ofbattery power. To reduce the power consumption, LTE introduces mecha-nisms for DRX by configuring a DRX cycle in the terminal which allows theterminal to monitor the control signaling only for a predetermined durationof time and turning off the transmission circuitry otherwise, allowing for sav-ings in power consumption. HARQ retransmissions take place regardless ofthe DRX cycle [28]. By implementing DRX procedures, the network is ableto trade-off between scheduling flexibility and power performance.
Chapter 3
Performance Evaluation of
Small-cell Deployments
The current and growing expansion need of cellular networks represents achallenge as laying out and deploying new infrastructure is extremely expen-sive. Small-cells are believed to be a cost effective solution to expand thecoverage and capacity of LTE and post-LTE networks [4].
Small-cells are low-power wireless access points that operate in licensedspectrum, used outdoor to enhance coverage, or indoor for enterprise or in-home usage. The concept of small-cells include femtocells, picocells andmicrocells.
In the case of in-home usage, femtocells provide high quality, high speedcellular access. They are deployed by end-users and connected to the opera-tor network by a digital subscriber line (DSL), cable modem or optical fiberconnection [24]. Because of the unplanned nature of femtocell deployments,they can suffer from severe inter-cell interference with neighboring femtocellsin dense deployments [26,54,68]. In addition, coordination is hardly feasibledue to delays induced by the backhaul infrastructure of these home femtocellnetworks.
In this chapter, we study a decentralized interference mitigation schemethat combines IR-HARQ with an interference cancellation decoder. Our per-formance evaluation based on analytical modeling and Monte Carlo experi-ments shows that our scheme is effective at combating interference withoutrequiring any coordination.
15
16 Chapter 3 Performance Evaluation of Small-cell Deployments
3.1 Interference in Femtocell Deployments
We concentrate on LTE and LTE-advanced technologies (so-called 4G) withOFDMA physical layers and explore alternative strategies to mitigate in-terference. OFDMA ensures orthogonality of the subcarriers and therefore,there is no intra-cell interference. However, interference can be experiencedfrom users in adjacent cells.
In LTE, HARQ is used to reduce errors in transmissions by retransmit-ting and combining the information when a frame is received with errors [27].With IR, additional and new redundancy information is transmitted andcombined with information already transmitted offering a coding gain. Aframe is retransmitted until it is discarded or a maximum number of re-transmissions is reached. From an information-theoretic perspective, mutualinformation is accumulated over retransmissions, increasing the probabilityto decode [20].
The throughput of HARQ has been investigated for Gaussian input sig-nals [20] over a Gaussian channel with fading and in [70], rate adaptationis performed to maintain a fixed outage probability. In [29] the diversity-multiplexing-delay tradeoff has been studied for the Multiple-Input Multiple-Output (MIMO) ARQ channel.
Unlike the previous work, we take advantage of the non-Gaussian natureof interference in home femtocell deployments where there are typically onlyone or two strong dominant interferers [60]. We consider signals comingfrom discrete alphabets in order to be able to benefit from the structure ofthe interference. Gaussian signals achieve the maximum spectral efficiency.However, practical systems make use of small, finite-size input alphabets.
Under these assumptions, we propose a decentralized strategy that com-bines interference cancellation decoding [32] with an incremental redundancyHARQ policy [20].
To evaluate the performance of this strategy, we develop an analyticalmodel of the throughput achieved by an HARQ protocol with an interferencecancellation decoder. In particular, our model builds on an information-theoretic characterization of the achievable rate with interference cancella-tion.
3.1.1 System Model and Assumptions
We focus on a downlink scenario. Without loss of generality, we currentlyconsider single antenna transmission. Furthermore, transmissions are slottedand perfectly synchronized. We have Nu transmitters, where node 0 is thetransmitter of interest and the remaining Nu−1 transmitters are interferers.We let dk be the distance between the node k and the receiver. 4G systemsare based on OFDMA physical layer [61]. We let y[m] be the received signalin a particular RB at time m. In LTE, every RB is defined as a group of K
3.1 Interference in Femtocell Deployments 17
subcarriers. Within a given cell, we assume that RBs are orthogonal to eachother. Hence, we can write the received signal y[m] at time m as
y[m]=K−1∑
j=0
Nu−1∑
k=0
√
Pkd−αk hk,j [m]µk,j [m]xk,j [m] + zj [m]. (3.1)
where xk,j [m] is the transmitted signal from node k in the jth subcarrier ofa particular RB, µk,j [m] is a so-called activity factor, zj [m] is thermal noise,Pk is the transmission power, α is the path loss exponent and hk,j [m] isthe channel coefficient. We model zj [m] as an independently and identicallydistributed (iid) zero-mean Additive White Gaussian Noise (AWGN) processwith variance σ2. If we concentrate on a particular subcarrier, the receivedsignal in the jth subcarrier at time m is
yj [m] =
Nu−1∑
k=0
√
Pkd−αk hk,j [m]µk,jxk,j [m] + zj [m] (3.2)
The random variable hk,j [m] is iid for each slot with a Rayleigh distri-bution. Hence, the channel coefficient remains constant during the durationof a slot. The activity factor models the traffic load and/or DiscontinuousTransmission (DTX) features of LTE systems [61]. We model µk,j with aniid Bernoulli distribution with parameter p. These features are taken intoaccount as they have a direct effect on the interference distribution.
The retransmission protocol is an HARQ scheme using IR [20]. For com-parison purpose, we also consider a simple ARQ scheme that retransmits thesame data block in case of unsuccessful transmission. The parameter Mmax
is the maximum number of ARQ rounds. Therefore, a given frame can beretransmitted at most Mmax times and is discarded if Mmax is reached. Weassume perfect Channel State Information (CSI) of the desired and interfer-ence signals at the receiver and we let R define the transmission rate givenby a particular MCS.
In the next section, we present the throughput analysis for our HARQprotocol with an interference cancellation receiver.
3.1.2 Average Throughput Analysis of HARQ with Interfer-ence Cancellation
Without loss of generality, our analysis can consider a single subcarrier andwe can drop the index j from (3.2). We consider a slotted (i.e. discrete-time)system where each slot corresponds to a frame transmission. We define thefollowing symbols:
• R̄ is the average throughput expressed in frames per second.
• p[m] is the probability that a frame is successfully decoded at slot m.
18 Chapter 3 Performance Evaluation of Small-cell Deployments
0
1
2
1
p[m]
q[m]
p[m + 1]
q[m + 1]
q[m + . . .]
p[m + 2]
p[m + Mmax]
q[m + Mmax]
Mmax
Mmax + 1
Figure 3.1: Retransmission Markov chain Xn: a frame transmissionattempt always initiates and finishes in state 0. A frame retransmission
corresponds to a transition from state i to i+ 1. A successful frametransmission corresponds to a transition from any state i = 0, . . . ,Mmax to
the state 0. Finally, the frame is dropped if state Mmax + 1 is reached.Note that q[m] = 1− p[m], ∀m.
The behavior of our system can be modeled by a discrete-time Markovchain [18]. However, because (1) a new channel coefficient hk,j [m] is pos-sibly drawn at every slot, and (2) the activity factor can change the numberof active sources at every slot, the Markov chain is inhomogeneous i.e. thestate transition probabilities can change over time.
Remember from Section 3.1.1 that a frame can be retransmitted at mostMmax times before being discarded. Hence, let Xn be the retransmissionstate of the source (see figure 3.1). The Markov chain Xn has Mmax + 2states (numbered from 0 to Mmax+1): a frame transmission attempt alwaysinitiates and finishes in state 0. A frame retransmission corresponds to atransition from state i to i+1. A successful frame transmission correspondsto a transition from any state i = 0, . . . ,Mmax to the state 0. Finally, theframe is dropped if state Mmax + 1 is reached. The transition probabilities
where pX(i, j) = Pr(Xn+1 = j|Xn = i). Each new frame transmission at-tempt1 corresponds to a trip on the chain Xn starting in state 0 and returningback to the state 0. The average throughput R̄ can be computed by dividingthe average number of successful frame transmissions per trip by the averageduration of a trip. For a trip from state 0 back to state 0, we can define tworandom variables:
• Ns is the number of successful frame transmissions per trip. Observethat Ns is equal to 0 or 1.
• T is the duration of a trip.
Now, following the approach in [51], we can write
R̄ =E0(Ns)
E0(T )(3.4)
where E0 is a Palm expectation [12] (see [51, eq. 9] for a definition specific
to the context of (3.4)). We do not have a closed-form expression for E0(Ns)and E
0(T ), but they can be evaluated by simulation. However, we need tobe able to compute the transition probability p[m + i] from state i to state0 at time m+ i.
We take an information-theoretic approach. Namely, without any inter-ference cancellation, the transition probability is computed as a function ofthe instantaneous mutual information conditioned on the channel knowledge.For interference cancellation, we follow the mechanism and modeling of [32]and modify the mutual information expression appropriately. Let Im denotethe mutual information between the received and the desired signal at timem and let R denote the target operating rate. In the following, the condi-tioning on the channel realizations and number of active users is implicitlyassumed. For IR-HARQ, it is shown in [20] that the mutual information isaccumulated over retransmissions when IR is used. Following [70], we have
p[m+ i] = Pr
(
m+i∑
k=m
Ik > R∣
∣
∣
m+i−1∑
k=m
Ik
)
(3.5)
for 0 < i ≤Mmax and p[m] = Pr (Im > R) , for i = 0. For ARQ, we have
p[m+ i] = Pr(Im+i > R) (3.6)
1As opposed to a retransmission, which is one transition on the chain.
20 Chapter 3 Performance Evaluation of Small-cell Deployments
for 0 ≤ i ≤Mmax. To evaluate equations (3.5) and (3.6) we need to computethe mutual information between node 0 and the receiver. To model the effectof interference cancellation, we follow [32]. For a receiver that cancels twointerferers, the mutual information is (we ignore the conditioning on thechannel for simplicity)
I(Y ;X0|X1, X2) = log S0 −1
S0S1S2
∑
x0
∑
x1
∑
x2
∫
y
p(y|x0)
× log
∑
x′
2
∑
x′
1
∑
x′
0p(y|x′0, x
′
1, x′
2)∑
x′
2
∑
x′
1p(y|x0, x′1, x
′
2)dy.
where Y is the received signal, X0 is the desired signal, X1 and X2 are theinterference signals, and Si, i = {0, 1, 2} is the size of the constellation forXi. If the structure of the interference is known, i.e. the constellation, thenthe receiver can use it to cancel the interferers. Note the discrete inputdistribution used to more accurately model practical systems. Also, to takethe remaining transmitters with index 3 to Nu−1 into account, we perform aGaussian approximation. With a receiver that does not cancel interference,we simply compute I(Y ;X0).
To visualize the mutual information with discrete signals, we plot in fig-ure 3.2 the Cumulative Distribution Function (CDF) of the mutual infor-mation for the case when there is no interference and we compare it withthe case of canceling two strong interferers. When interference is not takeninto account, it is clear that higher modulation orders will translate intohigher mutual information, but in an scenario with interference, there is notalways a clear gain from using higher modulation orders, it will depend onthe interference strength.
3.1.3 Performance of the HARQ Protocol in Femtocell De-ployments with Interference
In this section, first we give results for a comparison between ARQ andHARQ. For a different number of retransmissions, we obtain the throughputwith HARQ for two cases: without interference and with two strong inter-ferers. Then, we study a Manhattan-like topology when the two strongestinterferers can be canceled. Finally, we evaluate a scenario where interfer-ence is randomized by using an activity factor. This can be understood asthe re-use factor of the network.
All results are obtained evaluating the throughput expressions in Sec-tion 3.1.2 by Monte Carlo experiments averaged over fading and noise dis-tributions. The target rate R remains fixed for each simulation.
figure 3.3 compares the average throughput of the two ARQ protocols.We study a scenario with different modulations (BPSK, QPSK, 16QAM
3.1 Interference in Femtocell Deployments 21
0 1 2 3 4 5 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
I(Y;X)
F(x
)
64QAM
16QAM
QPSK
BPSK
(a) Mutual information CDF.
0 1 2 3 4 5 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
I(Y;X)
F(x
)
64QAM
16QAM
QPSK
BPSK
(b) Mutual information CDF, one interferer.
0 1 2 3 4 5 60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
I(Y;X)
F(x
)
64QAM
16QAM
QPSK
BPSK
(c) Mutual information CDF, two interferers.
Figure 3.2: In (a) we have the CDF of the mutual information underRayleigh fading and without any interference, (b) shows the CDF of themutual information with one interferer, and (c) for two interferers. The
interferers have the same power as the user of interest.
and 64QAM), without interference, and we set Mmax to 3. The Signal-to-Noise Ratio (SNR) goes from −10 to 25 dB and the target rate is R = 2bits/second. We clearly observe an improvement in throughput with HARQin comparison with the simple ARQ protocol without combining at the re-ceiver.
We can also see that it is possible to identify SNR intervals for eachmodulation order and that generally, at high SNR, 64QAM gives the highestthroughput. In figure 3.4, we plot the average throughput for different valuesof maximum retransmissions Mmax (in our terminology, retransmission isequivalent to round). We focus on QPSK modulation, HARQ and Mmax =1, 2, 3, 5, 10. We consider both no interference and two strong interfererscancellation.
When there is no interference (figure 3.4(a)), at most two rounds insteadof one gives a gain of around four times at a 0 dB SNR. However, goingfrom two rounds to three rounds gives a gain of only 0.1 in throughput.
22 Chapter 3 Performance Evaluation of Small-cell Deployments
−10 −5 0 5 10 15 20 250
0.5
1
1.5
2
2.5
3
3.5
4
4.5
SNR [dB]
Th
rou
gh
pu
t
HARQ 64QAM
HARQ 16QAM
HARQ QPSK
HARQ BPSK
ARQ 64QAM
ARQ 16QAM
ARQ QPSK
ARQ BPSK
Figure 3.3: We consider different modulations, for HARQ and ARQ, withMmax = 3 retransmissions. We show the average throughput for bothretransmission protocols and we see the improvement in throughput of
HARQ over ARQ for all modulations.
The later means that just increasing the maximum number of rounds is notsufficient to get a significant gain in throughput. Now, on figure 3.4(b) isa scenario with interference. We can observe that in the low SNR region,more retransmission rounds gives the highest throughput, however, at higherSNR, having only one retransmission gives the highest throughput. This is incontradiction of what is expected from having more opportunities to get theinformation decoded correctly. The reason for this result is that we fix thetarget rate for each case, which happens when there is no rate adaptation.This behavior suggests that the rate must be adapted depending on theinstantaneous SNR.
Next we present our results when combining IR-HARQ with interferencecancellation of the two strongest interferers. The topology is the Manhattan-like scenario in figure 3.5 and is a three by three grid topology, of size 30 by30 meters, with node 0 in the middle. The receiver is located 5 meters awayfrom the transmitter. This is a typical residential scenario.
In figure 3.6, we show the results for this topology. We consider QPSKmodulation, a target rate of R = 2 bits/second, maximum number of re-transmissions Mmax = 3 for both HARQ and ARQ. We show the aver-age throughput with and without interference cancellation for an SNR from−10 to 25 dB. For interference cancellation, two interferers can be canceled.We perform a Gaussian approximation on interferers that are not canceled.Hence, the SNR without interference cancellation is around −7 dB and theSNR if the two strongest interferers are decoded is around 6 dB. The respec-tive average throughputs are of 0.15 and 1. We observe a gain of around tentimes in throughput from canceling only the two strongest interferers.
3.1 Interference in Femtocell Deployments 23
−10 −5 0 5 10 15 20 250
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR [dB]
Th
rou
gh
pu
t
HARQ Mmax
=10
HARQ Mmax
=5
HARQ Mmax
=3
HARQ Mmax
=2
HARQ Mmax
=1
(a) Different number of ARQ rounds without interference
−10 −5 0 5 10 15 20 250
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
SNR [dB]
Th
rou
gh
pu
t
HARQ Mmax
=10
HARQ Mmax
=5
HARQ Mmax
=2
HARQ Mmax
=1
(b) Different number of ARQ rounds with two strong interferers.
Figure 3.4: We consider QPSK modulation, for HARQ, and we show theaverage throughput for different numbers of retransmissions
Mmax = 1, 2, 3, 5, 10. In (a) we plot the case without any interference andin (b) we see a different trend for the case with two strong interferers, (b)
shows that adapting the rate is an optimal way to increase the throughput.
24 Chapter 3 Performance Evaluation of Small-cell Deployments
Figure 3.5: Manhattan-like topology with the user of interest at the edge ofthe apartment and the rest of the interfering femtocells placed at the
middle of the surrounding apartments
−10 −5 0 5 10 15 20 250
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR [dB]
Th
rou
gh
pu
t
HARQ no interf
ARQ no interf
HARQ interf cancel
ARQ interf cancel
Figure 3.6: We look at the scenario in figure 3.5 and we consider QPSKmodulation, R = 2, Mmax = 3 for HARQ and ARQ. We show the average
throughput with no interference and with interference canceling twointerferers for an SNR from −10 to 25 dB. We do a Gaussian
approximation for the interferers that are not canceled. The SNR withoutinterference canceling is −7 dB and 6 dB for interference cancellation of the
two strongest interferes. The corresponding throughputs are 0.15 and 1,therefore there is a gain of ten times in throughput.
3.1 Interference in Femtocell Deployments 25
Finally, we investigate a bursty interference scenario where the interfer-ence is not constant. It is important to investigate how the throughput isaffected if we randomize the interference. In femtocell networks, being de-ployed by end-users, network and frequency planning becomes complicated,since the placement of the femtocells will be unknown for the operator. More-over, the users may be able to turn them on and off. In our model, if welet the nodes to be either active or inactive, this becomes equivalent to hav-ing a re-use factor in the network which creates an interference process thatis no longer ergodic. We investigated the case of interferers being present50% and 75% of the time, this is equivalent to setting the activity factorµ = {0.5, 0.75} and we compare it against an activity factor µ = 1. Thereare two interferers and we assume that all nodes transmit with unit powerfor simplicity. The SNR is defined then by SNR = 1
N0+2
−10 −5 0 5 10 15 20 250
0.5
1
1.5
SNR [dB]
Th
rou
gh
pu
t
HARQ µ=1
HARQ µ=0.75
HARQ µ=0.5
Figure 3.7: We consider QPSK modulation, we show the averagethroughput when the interference is randomized with an activity factor
µ = 0.5, 0.75, which means that the interferers will be active either half or75% of the time and we compare with the case of the interference presentall the time. We see that activity factor of 0.5 has the lowest throughputand that activity factor of 75% is closer to the corresponding curve for
µ = 1.
In figure 3.7 we show the curves for activity factors 0.5, 0.75 and 1 for thecase of canceling the two strongest interferers. We let all the users to have anactivity factor of µ = {0.5, 0.75, 1}. It means that both the user of interestand the interferers are not active all the time. This is equivalent to having are-use factor of 2 and 3/4 in time and frequency. If we look at figure 3.7, thethroughput when we use these values for the re-use factor is less than thecorresponding throughput with activity factor of 1. This result tells us that
26 Chapter 3 Performance Evaluation of Small-cell Deployments
under this scenario, it is better to use the whole bandwidth all the time. Itis important to highlight the fact that in this scenario, we consider a receiverthat can cancel the strong interferers. If we have such a receiver, then it isoptimal to use all resources or have a re-use factor of one. Having a re-usefactor of one also means that the spectral efficiency is higher.
Chapter 4
Mutual Information Analysis
of Interference Networks
LTE performance in terms of spectral efficiency and available data ratesis, relatively speaking, more limited by interference from adjacent cells ascompared to previous communication standards [27]. Means to reduce orcontrol the inter-cell interference can potentially provide substantial benefitsto LTE performance, especially in terms of the QoS provided to every user.
Efficient resource allocation algorithms are one of the key features inproviding high spectrum efficiency in LTE [34]. An intelligent schedulingalgorithm can, at the same time, help to reduce the impact of interference.
The performance of any particular scheduling algorithm can be evaluatedwith the help of exhaustive simulations, but it is time-consuming and has ahigh computational cost. By using information theory, we can analyze theachievable throughput, or spectral efficiency, under different system modelassumptions.
Rather than performing extensive simulations, we take an informationtheoretic approach to derive analytical expressions that represent the long-term throughput of the network and consider practical cases where there is aconstraint on the outage probability representing the latency of the protocol.We consider signals from discrete alphabets to model practical systems andwe consider cases where CQI feedback is either unavailable, or outdated.
4.1 Key Challenging Applications
In this section, we talk about some of the key emerging applications to whichour information theoretic analysis can be applied.
27
28 Chapter 4 Mutual Information Analysis of Interference Networks
4.1.1 Heterogeneous Networks
Cellular wireless networks can be divided into homogeneous and HetNets.Heterogeneous implies that there are different types of cells in the network,i.e. low-power base stations are distributed throughout a macro cell network.These low power base stations can be microcells, picocells, relays, femtocellsor distributed antenna systems [33]. On the one hand, microcells, picocellsand relays are deployed by the operator to increase the capacity and coveragein public places, enterprises buildings, etc. On the other hand, femtocells areuser-deployed at home to improve capacity. We generally denote low-powerbase stations by small-cells.
Cellular HetNets typically operate on licensed spectrum owned by thenetwork operator. The most severe interference is experienced when thesmall-cells are deployed on the same frequency carrier as the macrocells [47].More challenging interference scenarios are identified since interference cancome across layers (macro–small-cell, small–macrocell), for example, a macro-cell user far from the base station is transmitting at a very high power hurtingthe femtocells in the vicinity. Interference can also be experienced betweensmall cells in both the UL and DL channels (see figure 4.1). In the caseof inter-layer interference, the macrocell scheduler has to take into accountthe bursty interference from the femtocells since they will be serving only acouple of users.
Given the fact that in HetNets there is no controller managing the allo-cation process [11], operators will not be able to handle interference betweensmall-cells in a centralized manner (centralized frequency planning). Thedesign of distributed algorithms and techniques allowing for an efficient uti-lization of infrastructure is one of the key challenges in HetNets [44]. In thistype of approach, the small-cell adapts its performance independently fromother cells avoiding the need for any a priori centralized frequency plan-ning and without having to exchange information or sending it to a centralcontroller. They rely only on feedback, avoiding uncontrolled delays.
Link adaptation is one of the key features of HetNets. For instance, foran OFDM system, the allocation of subcarriers can be tailored according tothe interference conditions. However, mutual interference should be carefullyconsidered when designing heterogeneous systems.
4.1.2 M2M and Sparse Latency-Constrained Traffic
High-performance online gaming, M2M and sensor data communications areemerging massive applications for cellular networks. A typical example ofM2M applications in mobile environments is sensors connected to publictransport vehicles, to trains or to equipment in factories. It is predictedthat these applications, in addition to voice and Internet traffic, will be anintegral part of the traffic transported by LTE [61] and LTE-Advanced [57]
4.1 Key Challenging Applications 29
(a) Interference scenarios DL.
(b) Interference scenarios UL.
Figure 4.1: Figure (a) shows the interference scenarios for HetNets in theDL, figure (b) shows the interference scenarios for HetNets in the UL
networks. M2M is expected to account for a considerable amount of thetraffic of such networks [25].
M2M communications, part of the Internet of Things (IoT) revolution,is expected to create an increasing number of connected devices, whichwill exceed human-to-human communications over the following years (50billions machines against seven billion people for 2011) [25, 53]. A largeclass of the traffic generated by these emerging applications can require low-latency [41,71]. For example, for online gaming applications, the low-latencyis critical to offer the best game experience as possible [50]. Large portionsof M2M applications are expected to produce sparse traffic with low-delayconstraints. The two main reasons being power reduction through DRX andtransport of small sporadic packets to M2M devices. Concretely, a User-Equipment (UE) terminal emerging from an idle state to deliver a smallpacket to the network should reconnect for the smallest amount of time pos-sible to conserve power. Moreover, the paradigm of many sparsely connectedUEs transmitting sporadic traffic poses interesting problems related to re-
30 Chapter 4 Mutual Information Analysis of Interference Networks
Figure 4.2: Sparse traffic in a delay-constrained scenario. Traffic arrivals inthe eNB MAC layer are sparse as depicted in blue (there are three ofthem). The latency constraint is four slots, i.e. there are up to four
possible PDSCH channel allocations. Because of the sparse traffic, CQI isoutdated or unavailable on the first slot.
source allocation policies in a scheduled-access MAC protocol such as thatof 3GPP LTE.
In sparse and latency-constrained traffic scenarios, packet arrivals aresporadic and must be scheduled under a latency constraint (see figure 4.2).In this context, CQI is typically outdated or unavailable. Note that outdatedCQI also occurs because of moderate to high mobility, of insufficient uplinkCQI periodicity or of non-stationary inter-cell interference. The latter willbecome more and more important with LTE Release 10 networks and theirinherent heterogeneity. Hence, the scheduler must operate blindly for AMCand can only benefit from feedback after the first HARQ transmission roundin the form of ACK/NACK signaling.
The extremely large number of devices connected to the network, theexpected reliability of the service regardless of the operation environment,and low-latency requirements from applications such as emergency messages,video surveillance or health-care will require some enhancements to thenetwork including link adaptation protocols, modulation and coding, andHARQ schemes [48,69]. All these network optimizations will be included asa part of the LTE-Advanced standard since M2M communication is one ofthe main focuses in LTE-Advanced [25].
4.2 Related Work
Extensive research has explored variable-rate adaptation techniques. How-ever, very little attention has been paid to the more performance-limited caseof interference. Early work in [35] suggests a gain from adaptive policies. Byderiving the Shannon capacity regions of variable rate and power, it is shown
4.2 Related Work 31
that the maximum capacity is achieved when the rate is varied based on thechannel variations. More recently, [55] explores rate adaptation with suc-cessive interference cancellation receivers for MIMO systems with outdatedchannel state information and Gaussian signals. When no outage constraintis considered, [13] presents a Signal-to-Interference-plus-Noise-Ratio (SINR)threshold-based adaptation without the use of HARQ and in [66], the rate isadapted to the fading conditions under the concept of maintaining a certain“fairness” between users.
The throughput of HARQ has been investigated for Gaussian input sig-nals [20] over a Gaussian channel with fading and in the limit of infinite blocklength. In [67], the long-term throughput analysis of a HARQ protocol underslow-fading channels is presented for fixed-rate, variable-power transmissionsunder the framework of the renewal-reward theory of [20]. Rate adaptationfor HARQ protocols under delay constraints is studied in [70], and for time-correlated channels in [36] and [43]. In [16], rate and transmit power areadapted under perfect CSI. Power adaptation is also presented in [19] tominimize the outage probability and in [17], both power and rate control arederived through dynamic programming without outage constraints. Com-bined power and rate adaptation is also presented in [22], and in [23], theoptimization of either the packet drop probability or the average transmitpower is shown for the case of IR-HARQ with a maximum number of re-transmissions. In [64], the information-theoretic approach of [20] is adaptedto variable rate transmissions in the case of HARQ with IR. In [39], withoutconsidering HARQ, a mathematical framework based on a sum-rate analysisfor heterogeneous networks with partial feedback is developed.
The idea of changing the MCS for retransmissions is presented in [31],for IP video surveillance camera traffic by assigning additional redundancyto the retransmissions and reducing the estimated CQI.
Recent so-called rateless or fountain coding techniques with IR for additive-noise channels are also reported in [30], [58], [14], [37]. When combined witha HARQ link-layer protocol, these coding schemes allow for transmissionover unknown channels without the need for sophisticated rate adaptationpolicies, and whose instantaneous rate (or spectral-efficiency) depends onthe time the decoder is able to decode the message. The basic principle forthis type of transmission was introduced for content distribution over theinternet and broadcast networks by Luby [49] using so-called LT-codes forerasure channels. These were improved by Shokrollahi with his invention ofRaptor codes [63]. The latter were then adapted for AWGN channels in [56].
All of these coding strategies are structured, and, in particular Perry etal’s Spinal Codes, can approach Shannon’s AWGN channel capacity withvarying degrees of encoding and decoding complexity provided the numberof transmissions is allowed to grow without bound. Although not shownin [30] [14] it may very well be true for any ergodic time-varying additive-
32 Chapter 4 Mutual Information Analysis of Interference Networks
noise channel. An extension of the promising superposition coding techniquedesigned for successive decoding at the receiver considered by Erez et alwas also described for time-varying channels without an a priori stochasticmodel [30]. This considered the performance of their rateless coding con-struction for a small number of transmission rounds.
In this work, we consider similar rateless strategies for time-varying chan-nels for a finite and small number of transmission rounds, potentially allowingfor a residual outage probability after the maximum number of rounds. Im-posing a quasi-finite duration for transmission is often required to minimizelatency in data transmission networks. For instance, the HARQ protocol ofLTE reference channels [8] is tuned to offer an approximate 1% outage rateafter two transmission rounds which allows for a one-way latency of 10ms for99% of transmissions. This can, of course, be tuned to offer different latency-throughput tradeoffs. Since the maximum number of transmission roundsis fixed in such protocols, it seems natural that the number of dimensionsused in each round should be optimized in order to maximize throughput,by progressively decreasing code rate across rounds. We should note thatthe latter is not a requirement in rateless coding with an unbounded numberof transmission rounds.
4.3 Initial Analysis for Interference-free Networks
We start our analysis by looking into interference-free networks, i.e. we donot consider interference created by neighboring transmitters and we focuson single antenna systems, Single-Input Single-Output (SISO), although ourmodel can be extended to MIMO.
4.3.1 Signal Model and Assumptions
In the following, we present the signal model and assumptions for this sec-tion. Without loss of generality, we consider OFDM signaling. The ULof an LTE system uses SC-FDMA modulation. Our joint HARQ and rateadaptation policy applies equally, but the signaling details differ. Therefore,for a particular subcarrier j, let x denote the complex-valued transmittedsymbol, z denote the additive white Gaussian noise (AWGN), and h denotethe channel gain. Both z and h are modeled with a zero-mean and unitvariance complex Gaussian random variable. With a total of K subcarriers,the received signal (yj), in a particular subcarrier is
yj = hjxj + zj , j = 1, 2, . . . ,K. (4.1)
We consider a block-stationary Rayleigh fading channel model. Fading re-mains static for the duration of a HARQ round but varies between retrans-missions. The HARQ feedback channel is assumed to be error-free. CQI
4.3 Initial Analysis for Interference-free Networks 33
can be received after each round. However, prior to the first round, CQImay or may not be available. As explained earlier, this can occur becauseof sparse traffic. Furthermore, because of fast-fading, (low) mobility or non-stationary inter-cell interference, CQI can at any round be simply unusable.Consequently, we analyze cases where, at the first transmission round, CQIis either outdated or simply unavailable. On the further rounds, we keep onassuming that CQI is not available. But, we take advantage of the one bit ofACK/NACK information given to the transmitter after each HARQ round.
For the outdated CQI case, it is assumed that the fading statistics areavailable to the transmitter. This assumption is reasonable because theeNB scheduler can maintain a database of channel measurements in its cell,allowing it to derive the fading statistics over time.
4.3.2 Modeling and Optimization of a Resource SchedulingPolicy
We consider a one-shot transmission model where one transport-block ofsize NTB arrives in sub-frame n and must be served at maximum spectral-efficiency under a latency constraint. We denote by Mmax the maximumnumber of transmission rounds. To characterize code performance and theeffect of the channel, we use the instantaneous mutual information in eachtransmission round. This theoretic measure, although asymptotic, providesa very accurate indication of potential performance in LTE, whose coded-modulation subsystem performs close to asymptotic limits. Let Hr denotethe vector of channel realizations in the fth transmission round. Then I (Hr)denotes the corresponding instantaneous mutual information. Accordingly,
I (H1, . . . , HMmax)
defines the mutual information accumulated over Mmax transmission rounds.In order to compute the mutual information, we assume Gaussian inputsignals (upper-bound on QAM modulation). For example, let us considerone subcarrier of a SISO system without interference and let P denote thereceived power, hr is the channel response at round r and N0 is the noisepower, then
I(H1, . . . , HMmax) =
Mmax∑
r=1
log2
(
1 +P |hr|2N0
)
. (4.2)
Generalizing the notation from [20], the probability of decoding a transport-
34 Chapter 4 Mutual Information Analysis of Interference Networks
block in round r with Nj as the number of dimensions used in round j is
Pr(
I(H1, · · · , Hr) > Rr
r∑
j=1
Nj ,
I(H1, · · · , Hn) < Rn
n∑
j=1
Nj , ∀n < r)
. (4.3)
Let Pout,n denote the target transport-block error probability after n trans-mission rounds. The latency constraint is expressed by ensuring that theprobability that the transport-block is not served after Mmax transmissionrounds is below Pout,Mmax
. Under this framework, rate adaptation is theoptimization of the rate sequences Rr such that (1) the packet error proba-bility remains below Pout,Mmax
after Mmax transmission rounds and (2) thespectral-efficiency is maximized. The optimization is carried out as a func-tion of the distribution of I (H1, . . . , HMmax
).For simplicity, we consider at most two retransmission rounds (ARQ
rounds), but our policy can also be applied for more than two. We considerthree scenarios
1. Minimal-latency: a trivial case of serving the packet in one round whichcorresponds to the minimal-latency rate adaptation policy.
2. Latency-constrained with no prior CQI: we consider two transmissionrounds and no information about the channel.
3. Latency-constrained with outdated CQI: we consider again two trans-mission rounds, but unlike the previous case, we assume that we haveoutdated information about the channel with some correlation withthe actual channel.
For simplicity and in the interest of obtaining semi-analytical results, weconcentrate on one subcarrier, i.e. that Hr is a scalar. In Chapter 5, usinga fully-compliant LTE modem implementation, this is reconsidered using amultipath 3GPP Extended Pedestrian (EPA) channel model.
Scenario Analysis: Minimal-latency
We consider first the trivial case of serving the transport-block in one round.This is the minimal-latency rate adaptation policy. The rate allocation lawfor R1 is given by the solution to
Pr(
I(H1) < R1
)
= Pout,1. (4.4)
Without any a priori information regarding the channel statistics, this es-sentially says that the best that can be done is to transmit with the lowest
4.3 Initial Analysis for Interference-free Networks 35
spectral-efficiency coding scheme (i.e. lowest MCS) to minimize latency.With a priori information, the largest MCS such that the probability ofchannel realizations requiring a smaller MCS is still below the threshold ischosen.
Let Hout denote the channel corresponding to outdated CQI. If stale CQIis available prior to transmission of the transport-block, then the rate shouldbe chosen such that
Pr (I(H1) < R1|Hout) = Pout,1 (4.5)
in order to take into account the outdated CQI.
Scenario Analysis: Latency-constrained with no Prior CQI
We now consider the case with two transmission rounds. Let B define thenumber of information bits to be transmitted. Let NT denote the totalnumber of dimensions available and let N1 denote the number of dimensionsused in the first round. Hence, the rate in the first round is R1 =
log2 BN1
, and
the rate in the second round R2 =log2 BNT
. We define
ρ =N1
NT
. (4.6)
and we can relate R1 to R2 with R2 = ρR1. Let R̄ denote the overall spectralefficiency. With Pout,1 as the outage probability after the first round, we have
R̄ = R1 (1− Pout,1) + Pout,1R2
= R1 (1− Pout,1) + Pout,1ρR1. (4.7)
We want to maximize R̄ such that the probability of outage after the sec-ond round is below the given constraint Pout,2 For the first round, there isno feedback information. The outage probability Pout,1 is unknown but itdepends on H1 and the SNR. We can relate R1 to Pout,1 as follows. Fromequation (4.4), we have
Pr (I(H1) < R1) = Pr(
log2(
1 + SNR|h1|2)
< R1
)
= Pout,1. (4.8)
Consequently, we obtain
R1 = log2
(
1− SNR ln(1− Pout,1))
. (4.9)
In the second round, feedback about the previous round is available. Theoutage probability is now given by
Pr(
I(H1, H2) < R2|I(H1) < R1
)
= Pout,2 (4.10)
36 Chapter 4 Mutual Information Analysis of Interference Networks
We can rewrite equation (4.10) as follows
Pr(
I(H1, H2) < R2|I(H1) < R1
)
=Pr(I(H1, H2) < R2, I(H1) < R1)
Pr(I(H1) < R1)
=
∫
2R1−1SNR
0 e−|h1|2d|h1|2Pout,1
−∫
2R1−1SNR
0 e−a−|h1|2d|h1|2Pout,1
=Pout,2 (4.11)
where
a =
(
(
2R1
1 + SNR|h1|2)
ρ
1−ρ 1
SNR
)
− 1
SNR(4.12)
and the limits stem from the fact that if I(H1) < R1 then |h1|2 < 2R1−1SNR .
The integrals in equation (4.11) are evaluated numerically.To find the optimal value of R1 in the first round, we perform an extensive
exploration on Pout,1, given that we want to maximize equation (4.7) andsubject to the constraint Pout,2 in equation (4.11).
Scenario Analysis: Latency-constrained with Outdated CQI
Because of the sparse traffic characteristic, of moderate to high mobility, ofinsufficient uplink CQI periodicity or of inter-cell interference, we investigatecases where the UL CQI is outdated or unavailable. In such cases, the sched-uler can only benefit from binary feedback after the first HARQ transmissionround (in the form of ACK/NACK signaling [61]).
We make the additional assumption that the channel remains constantover the two transmission rounds and let h = h1 = h2. Furthermore, wedenote by hout the channel value that corresponds to the outdated CQI. Inorder to model a possible correlation between hout and h, we use the followingmodel. Let λ be the correlation parameter, then
h =√λhout +
√1− λh
′
where hout and h′
are i.i.d. Gaussian-distributed random variables. Notethat in this case,
λ = E [houth∗] .
In addition, |h|2 is a non-central Chi-square random variable with two degreesof freedom. We follow the same general procedure to obtain the throughputand probability of outage than in the previous cases. However, the spectralefficiency is a function of the outdated CQI and we have to average over thedistribution of |hout|2.
First, let γ1 =√
2R1−1SNR be the outage threshold in the first round and
γ2 =√
2R2−1SNR be the outage threshold in the second round. Then, Pr (h > γ1)
4.3 Initial Analysis for Interference-free Networks 37
represents the probability of having a successful transmission in the firstround, Pr (h < γ1, h > γ2) is the probability of being unsuccessful in thefirst round but successful in the second round, and Pr(h < γ2) gives theprobability of being in outage. All these probabilities are a function of theCDF of |h|2. The non-centrality parameter of |h|2 is s2 = λ|hout|2. LetFχ (χ) denote the CDF of |h|2. It can be expressed in terms of the MarcumQ-function [59], i.e.
Fχ (χ) = 1−Q1(s, χ) (4.13)
where QM (a, b) is the Marcum Q-function with M degrees of freedom andparamters a and b.
The overall spectral efficiency R̄ over the two ARQ rounds can be writtenas
To find the optimal rates, we first obtain R2 from the outage constraintPout,2. Since we know that h < γ2 implies an outage, R2 is given by solvingequation (4.15) for R2. Therefore
Pout,2 = Pr
(
|h|2 < 2R2 − 1
SNR
)
= Fχ
(
γ22)
(4.15)
Next, to find the value of R1 that will maximize the overall spectral efficiency,we first write (4.14) in terms of the Marcum Q-function. We have
where a1 = a2 = s, b1 = γ1, and b2 = γ2. We now take the derivative ofequation (4.16) with respect to R1, and we solve for R1 when the derivativeis zero. We obtain
∂R̄
∂R1=
∂Q1(a1, b1)
∂R1R1 +Q1(a1, b1)−
∂Q1(a1, b1)
∂R1R2
=∂Q1(a1, b1)
∂R1(R2 −R1) +Q1(a1, b1) = 0. (4.17)
To find the derivative of the Marcum Q-function in (4.17), we used [10].
4.3.3 Numerical Results
In this section, we present numerical results in terms of (1) the probability ofoutage and (2) the achieved spectral efficiency. Remember that we assumeGaussian signaling to compute the mutual information. Throughout thissection, we fix the spectral efficiency to 2 bits per channel use. The maximumnumber of retransmissions is one round (at most two rounds).
Figure 4.3 presents the minimum SNR necessary to achieve a given outageprobability Pout,2. For a given value of Pout,2, we calculate the corresponding
38 Chapter 4 Mutual Information Analysis of Interference Networks
5 10 15 20 25 30 35
10−4
10−3
10−2
10−1
SNR [dB]
Po
ut,
2
outdated channel, λ=50%
outdated channel, λ=10%
uncorrelated channels
Pout,1
fixed to 50%
no feedback
Figure 4.3: For different values of the probability of outage after the secondround Pout,2, we calculate the corresponding SNR for the different
scenarios. The symbol λ is the correlation coefficient between the actualchannel and the channel corresponding to outdated CQI information. Wecompare a correlation coefficient value of 50%, 10% and uncorrelated case.
For comparison purposes, we also plot two more cases. First when noACK/NACK feedback is available from the HARQ process. Second, whenPout,1 is fixed to 50% with ρ = 0.5 to make sure that 50% of the dimensions
are used in each round.
4.4 Interference Networks Analysis 39
SNR for our rate adaptation policy. For the scenario (3) from Section 4.1.2,the outdated CQI has a correlation coefficient with the actual channel ofλ = 10% or λ = 50%. For comparison purposes, we consider two morecases in addition to the scenario (2) and (3) from Section 4.1.2. First weevaluate a case where we force the probability of outage after the first roundto 50%, fixing ρ = 0.5 to make sure that 50% of the dimensions are usedin each round. Typically, while conventional systems try to ensure a 10%outage probability per slot, we observe from our results that a higher valuegives, in fact, a higher overall spectral efficiency. Second, we evaluate acase where no feedback at all is available i.e. when we can not even receiveACK/NACK from the HARQ process. This highlights the significant gainfrom adapting the rate across rounds with only one bit of feedback, even inthe case without any CQI information. The gain is even higher when onlyoutdated CQI information is available. Our rate adaptation policy gives azero probability of outage without the need of having a high SNR. From theresults in figure 4.3, we can observe that it does not make a difference toincrease the SNR above 12.5 dB for the case without CQI. We show thatadjusting the dimensions used in each round results in almost causal feedbackperformance. In our scenarios, the two rates are simply controlled by ρ = R2
R1,
which depends on the SNR and target outage probability Pout,2. We onlyneed one bit of feedback, which we get causally from HARQ. In fact, it isthe state of the channel that chooses the code rate. By choosing the rate inthe first round as high as possible, we can guarantee a probability of outageafter the second round while maximizing the spectral efficiency.
Figure 4.4 presents the overall spectral efficiency obtained for a givenSNR. We set Pout,2 to 1%. For the outdated CQI case, we consider λ = 50%and λ = 10%. For reference purposes, we also plot the ergodic capacity(Rayleigh channel capacity), i.e. perfect rate adaptation. Finally, we con-sider a scenario where the rate in the first round is chosen as the one thatcorresponds to a probability of outage after the first round of 50%. Thisvalue is chosen because it gives the highest spectral efficiency. Fixing theprobability of outage after the first round to more or less than 50% gives, infact, a lower overall spectral efficiency.
From our results, we see a significant improvement in spectral efficiencyeven in the case without CQI. When we can benefit from outdated CQI, weachieve a performance close to the ergodic capacity. If we look at the resultsfor the ratio of dimensions for the second round (right axis and dashed lines),we can see that as the SNR gets higher, more dimensions are used in thesecond round.
4.4 Interference Networks Analysis
We now consider a network with interference from adjacent cells.
40 Chapter 4 Mutual Information Analysis of Interference Networks
0 5 10 15 20 25 300
1
2
3
4
5
6
7
8
9
SNR [dB]
Sp
ectr
al eff
Rayleigh channel capacity
outdated channel λ=50%
outdated channel λ=10%
uncorrelated channelsP
out,1 fixed at 50%
ρ
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
rati
o o
f d
imen
sio
ns
Figure 4.4: The axis on the left (solid lines) shows the spectral efficiencyversus SNR for the different scenarios. We set Pout,2 to 1%. The symbol λ
is the correlation coefficient between the actual channel and theoutdated/stale CQI information. We compare correlation coefficients of50%, 10% and uncorrelated case. For comparison purposes we plot the
curve for the ergodic capacity and Pout,1 fixed to 50% with ρ = 0.5 to makesure that 50% of the dimensions are used in each round. The axis on theright (dashed lines) shows the ratio of dimensions used in the two rounds.
4.4.1 Modeling and Assumptions
We consider a slotted transmission scheme and we take an information-theoretic approach to analyze the throughput performance. When thereis more than one user, we assume that all transmissions in every slot aresynchronized and we randomize the interference process with the use of ac-tivity factors. The latter models sporadic interference patterns characteristicof future heterogeneous networking deployments, in particular the interfer-ence seen from small-cell base stations with bursty traffic in the receiver of amacrocell user. They can also model dual-carrier networks with cross-carrierscheduling. In this type of network, we can talk about clean and dirty carri-ers. On the one hand, clean carriers are used by the macrocell to carry theirdata plus signaling for small-cells because of their controlled interferenceproperty. On the other hand, dirty carriers are interfering carriers where the“cleaning” is done with the use of HARQ.
We consider a maximum of Mmax HARQ transmission rounds and thechannel is iid or constant over all the transmission rounds of the protocol.After each transmission we receive an error-free acknowledgment (ACK orNACK) indicating a successful or unsuccessful transmission. We define theprobability of outage as being unsuccessful to correctly receive the informa-tion at the end of the HARQ protocol. This probability translates to the
4.4 Interference Networks Analysis 41
Figure 4.5: Coding Model
latency of the protocol and QoS in our system.In general, we define Rr as the code rate at the rth round. For a particular
user, we define the number of dimensions in time as Tdim and the numberof dimensions in frequency as Lr. Let L′r be the number of dimensions infrequency up to round r. Then, at each transmission round, the total numberof dimensions is L′rTdim. Assuming the channel does not vary during Tdim
time dimensions and for a packet length of B information bits, the rate Rr
at the rth round, in bits/dim is given by:
Rr =log2B
L′rTdim
bits/dim. (4.18)
In IR-HARQ, the retransmission consists of the same set of information bitsas the original, however, the set of coded bits are chosen differently and theymay contain additional parity bits. In each of the transmission rounds thereare LrTdim dimensions, however, this number is not necessarily the sameacross rounds according to the LTE standard [7] (see figure 4.5).
In the context of LTE, the number of physical dimensions LrTdim refersto the number of resource blocks allocated to one user in one subframe of 1ms duration, i.e. one Transmission Time Interval (TTI). There are at mosttwo transport blocks delivered to the physical layer in the case of spatialmultiplexing [28]. In a single-user LTE system, there is only one transportblock in one TTI, representing only one codeword “in the air” at the sametime. Each transport block is carried by an HARQ process, and each processis assigned to a subframe (number of processes is fixed). In our model, ifthe number of dimensions for a user is less than the maximum number ofavailable resources NT , (LrTdim < NT ), then the rest will not be utilized.Although not possible in the current LTE standard, one could propose toassign the unused resources to transmit multiple codewords in parallel (at thesame time), to increase the throughput. In a multiuser system, the remainingdimensions would be allocated to other users and thus the efficiency of theprotocol should be chosen to maximize the aggregate spectral efficiency ofthe cell.
42 Chapter 4 Mutual Information Analysis of Interference Networks
Let Psucc1 be the probability of having a successful transmission in thefirst round, and Psuccr,failr−1 the probability of not having a successful trans-mission in the (r−1)th round, but being successful in the rth round. Finally,let Pout represent the probability of outage at the end of the protocol. Theoverall throughput can thus be expressed as:
R̄ = Psucc1R1 +
Mmax∑
r=2
Psuccr,failr−1
(
Rr
r
)
bits/dim. (4.19)
where the outage probability is given by Pout = Pr(r = Mmax + 1) = 1 −∑
Mmax
r=1 Psuccr .In the case of an upper layer ARQ, the throughput expression becomes [20]:
R̄ = [1− Pout]
[
Psucc1R1 +
Mmax∑
r=2
Psuccr,failr−1
(
Rr
r
)
]
bits/dim. (4.20)
We consider Nu transmitters, where user 0 is the transmitter of interest,and the remaining Nu−1 transmitters are interferers. We model an OFDMAphysical layer with K subcarriers. We let µj,k be the activity factor, Pj thetransmission power, and xj,k the input signal of the jth user on the kthsubcarrier. We assume discrete signals with equal probabilities and size ofthe constellation |xj,k| = S, zk is the zero mean complex Gaussian noise withvariance σ2. Since we assume Rayleigh fading, hj,k is a circularly symmetriccomplex Gaussian random variable with unit mean. The received signal y isgiven by:
y =Nu−1∑
j=0
√
Pj
K−1∑
k=0
hj,kµj,kxj,k + zk (4.21)
Variations in the channel are caused at the receiver because of the activityfactor plus the frequency shifting from the resource allocation process. Forthe interfering users, the channel variation depends on whether we considerthe UL or DL. In the UL, it is caused by the interference coming from differ-ent user terminals. In the DL, the activity factors will introduce variationsoriginated from the fact that the interfering cells are not active the wholetime.
To model heterogeneous networks where the interference is not constant,we let the user of interest to be active all the time (macrocell user) andwe let the small interfering cells to transmit with a certain probability (seefigure 4.6).
In the next section, we provide a motivating example on the importanceof adapting the rate and physical dimensions across the rounds of the HARQprotocol.
4.4 Interference Networks Analysis 43
Figure 4.6: Downlink of a macrocell with a femtocell interfering.
4.4.2 Simple Interference Analysis in Zero-outage
For the sake of analytical tractability, we start by looking at the special caseof (4.21) where the hj,k are fixed (AWGN channel) and we assume Gaussiansignals. We consider one interferer and we model it with an activity factor,which means that the interferer could be active or inactive. The activityfactor is Bernoulli distributed with probability p.
The rate with Gaussian codebooks that can be achieved by the proto-col depends on the interference state (interference active or inactive). LetRH be the capacity that can be achieved without interference, and RL thecorresponding capacity with interference, which are given by:
RH = log2(1 + SNR1) (4.22)
RL = log2
(
1 +SNR1
1 + SNR2
)
(4.23)
where SNR1 is the SNR for the user of interest, SNR2 is the correspondingSNR for the interferer and we assume unitary noise variance.
We consider a HARQ protocol with two rounds and we define ρ = N1NT
,where N1 is the number of dimensions used in the first round and NT hasbeen previously defined as the total number of dimensions. Then for a packetof length B bits, the rate in the first round is R1 = 1
ρNTlog2B = RH and
in the second round R2 = 1NT
log2B = RL. Therefore, RL = ρRH , and
ρ = RL
RH.
In the remainder of this section, we derive the zero-outage throughputwith and without feedback and we consider also the case of a residual outageat the end of the protocol with feedback.
Zero-outage Throughput Without Feedback and no Delay Con-straint
We now look at the case of no feedback. Let the activity factor µ definethe state of the interference. We consider ON/OFF interference, therefore,if µ = 0 there is no interference and µ = 1 means that the interference is
44 Chapter 4 Mutual Information Analysis of Interference Networks
active and it happens with probability p. Then the throughput R̄ with zerooutage (without delay) is given by:
R̄ = EµI(X;Y |µ) (4.24)
= (1− p)RH + pRL (4.25)
It is interesting to note that (4.25) is the ergodic capacity (average overall possible states). In the next section we explore the case when feedbackbecomes available and we look at the case of more than two transmissionrounds.
Zero-outage Throughput with Feedback
In this case, we assume that we have feedback from the HARQ protocol andwe vary the tolerable latency by fixing the maximum number of transmissionrounds Mmax, but still assume zero-outage probability. Then, given thatwe want zero-outage at round Mmax, we choose the rate that guaranteessuccessful decoding (i.e. RL). We choose the rate in the first round to be ashigh as possible, and the intermediate rates are at the optimal value betweenRL and RH . Therefore, the rate at the rth round is given by:
R1 =log2B
ρ1NT
= RH r = 1 (4.26)
Rr =log2B
(∑r
j=1 ρj)NT
=
(
ρ1∑r
j=1 ρj
)
RH 2 ≤ r < Mmax (4.27)
RMmax=
log2B
NT
= RL = ρ1RH ⇒ ρ1 =RL
RH
r = Mmax (4.28)
In this case, the throughput expression for Mmax rounds is given by:
R̄ = (1−p)RH +
Mmax−1∑
r=2
pr−1(1−p)
(
ρ1∑r
j=1 ρj
)
RH +p(Mmax−1)RL (4.29)
For the rates to be achievable, we observe that there is a restriction on theratio of dimensions after the second round ρr, r > 1. This restriction comes
from the fact that the rate after round r is(
∑r−1j=1 ρj
)
NTRL + ρrNTRH
which means that after round r we decode if:
r∑
j=1
ρj
NTRr <
r−1∑
j=1
ρj
NTRL + ρrNTRH (4.30)
Rr =
(
ρ1∑r
j=1 ρj
)
RH <
(
∑r−1j=1 ρj
)
RL + ρrRH∑r
j=1 ρj
ρr > ρ1
1−r∑
j=1
ρj
(4.31)
4.4 Interference Networks Analysis 45
−10 −5 0 5 10 15 200
0.2
0.4
0.6
0.8
1
SNR [dB]
rati
o o
f d
ime
ns
ion
s
−10 −5 0 5 10 15 200
1
2
3
4
5
Th
rou
gh
pu
t (b
its/d
im)
ρ1
ρ2
4 rounds
3 rounds
2 rounds
1 round
Figure 4.7: The axis on the right (solid lines) shows the zero-outagethroughput for the HARQ protocol with different number of rounds, whilethe axis on the left (dashed lines) shows the ratio of dimensions per round
for the three rounds, zero-outage HARQ protocol. In both cases thechannel is AWGN with Gaussian signals and there is one interferer withprobability p = 0.5. The interference strength is the same as the user of
interest (SNR1 = SNR2).
If we look at figure 4.7, the solid lines and the right axis show thezero-outage throughput for the HARQ protocol with a maximum numberof rounds Mmax = {1, 2, 3, 4}. We can see that there is a high gain whengoing from one to two rounds and after three rounds there is only a marginalgain. The dashed lines and left axis show how the dimensions are beingdistributed across the rounds of the protocol. We illustrate the case of threerounds (i.e. Mmax = 3) and we look at the proportion of physical dimensionsused in each round (ρr). In both cases, the interference strength is the sameas the user of interest (SNR1 = SNR2), the channel is AWGN and we assumeGaussian signals with one interferer active with probability 50%. If we lookat 10 dB SNR, we observe that 20% of the dimensions are used in the firstround, 15% in the second round and the remaining 65% are left for the thirdround. In general, we observe that at high SNR, the number of dimensionsused in the first round decreases, leaving progressively more dimensions to
46 Chapter 4 Mutual Information Analysis of Interference Networks
the last round. If we think about the almost blank subframes (ABS) featureof LTE, which restricts the transmission in the cell if there is interference, wewould have a lower spectral efficiency than the average rate and therefore,by adapting the dimensions one can achieve a higher spectral efficiency evenin the presence of interference.
Throughput With Outage and Feedback
In this case, we allow the protocol to have a residual outage probability whichis overcome by an upper layer ARQ process on top of the IR-HARQ [46], [20],and we assume that we have feedback. For two rounds, from (4.20) thethroughput is given by:
R̄ =(1− Pout(ρ,R2))
[
(1− Pout,1(ρ,R2))R2
ρ+ Pout,1 (1− Pout,2(ρ,R2|out1))R2
]
(4.32)
where Pout,2(ρ,R2|out1) is the outage probability at the second round, giventhat there was an outage in the first round, and Pout,r is the probability ofoutage at round r.
I(µr) is the mutual information as a function of the state of the interfer-ence at round r, and it is defined by µr:
I(µr) =
{
log2(1 + SNR1) µr = 0
log2
(
1 + SNR1
1+SNR2
)
µr = 1(4.33)
Now, we can define the probabilities in (4.32) as follows: Pout,1(ρ,R2), theoutage probability at the first round, is given by:
Figure 4.8: Throughput of the two rounds HARQ protocol in an AWGNchannel with Gaussian signals. There is one interferer with probability
p = 0.05, 0.5.
where Pr((1− ρ)I(µ2) + ρI(1) < R2) =
p if I(1) < R2 < ρI(1) + (1− ρ)I(0)
0 if I(1) > R2
1 R2 > ρI(1) + (1− ρ)I(0)
(4.35)
Finally, Pout(ρ,R2) is the probability of outage after the second round,independently of the interference state at the first round and it is given by:
Pout(ρ,R2) =
{
p(1− p) R2 > ρI(1) + (1− ρ)I(0)
p2 R2 < ρI(1)(4.36)
Figure 4.8 shows the throughput of the HARQ protocol with two trans-mission rounds. There is one interferer with probability p = {0.05, 0.5}.We compare the zero-outage throughput against the throughput that allowsan outage at the end of the protocol. We also plot the maximum capac-ity achieved with one round and no interference RH and the correspondingcapacity for interference RL. If we look at the case of 50% probability ofinterference, we can see that the zero-outage throughput is higher for allSNR values, however, if we look at a case with a lower probability of havinginterference (p = 0.05, or 5%), we have almost the same throughput, exceptat high SNR, where the throughput with an outage is slightly higher. In thiscase, we also see that the capacity that can be achieved by adapting the rateand dimensions gets close to the capacity achieved without interference.
48 Chapter 4 Mutual Information Analysis of Interference Networks
Discussion
If we consider the case with the ergodic capacity and no feedback, we trans-mit NT dimensions per channel realization. Therefore, we have the averagecapacity:
Eµ =
{
log2(1 + SNR1) µ = 0
log2
(
1 + SNR1
1+SNR2
)
µ = 1(4.37)
where µ is the state of the interference. Now, if we consider a channelwith feedback of the state of the interference (non-causal feedback). Thenat round r, the transmit signal is a function of the message W and theinterference state µ:
{
xr = f(W,µ) r > 1
x1 = f(W )(4.38)
To get an insight into how a rate-adaptive scheme performs when changingthe number of dimensions across rounds, we focus on the case of the HARQprotocol with two transmission rounds. At round r, if µ = 1, then there isno transmission, and it happens with probability Pr(µr = 1) = p. However,if there is no interference, µr = 0, it transmits with NT
1−p dimensions, and
in this case we get a throughput= (1 − p)
(
log2(1+SNR1)NT
(1−p)
NT
)
= log2(1 +
SNR1) which is the maximum achievable spectral efficiency. When feedbackbecomes available, it allows the scheme to perform better than the ergodiccapacity. The latter is in contrast to the work in [20] where in the infinitedelay case, the authors conclude that the maximum that can be achieved isthe ergodic capacity. The difference comes from the fact that in [20] there isalways a fixed bandwidth allocation for each user, regardless of the state ofthe channel. In our case, we dynamically adapt the bandwidth for each userdepending on the interference conditions of past transmissions for the samecodeword. From the perspective of the scheduler, the bandwidth is betterdistributed.
From our initial analysis, we can conclude that the highest spectral effi-ciency that can be achieved happens in the case of the zero-outage protocolwhere increasing the delay becomes beneficial to a certain point and bringsonly a marginal gain after this point. In the next section, we look at practicalinterference scenarios where having zero-outage throughput is not possible.However, a constraint on the outage probability can be imposed.
4.5 Practical Interference Networks Analysis
To model practical systems, we derive expressions for the mutual informationassuming discrete constellations. We target LTE Release 10 networks with anOFDMA physical layer, and we study both the single-user and one dominant
4.5 Practical Interference Networks Analysis 49
interferer cases. For the sake of simplicity, we show the derivations focusingon one subcarrier, with unitary power and distance, and we drop the indexes.
Let Hr denote the vector of channel realizations in the rth round, thenIr(H) = Ir(Y ;X|H) denotes the corresponding instantaneous mutual in-formation at round r. For IR-HARQ, mutual information is accumulatedover retransmissions. In the case of bursty interference, this permits someaveraging of the fading and interference affecting the signal [65].
For a particular user, we define the mutual information at round r, inbits, as:
Ir(H) = Tdim
r∑
j=1
Lj∑
k=1
Ij,k(Hj) (4.39)
where Ij,k(Hj) is the mutual information for the user at round j and sub-carrier k, and it is given by
Ik(H) = Ik(Y ;X|H) =1
S1S2
∑
x1
∑
x2
∫
y
f(y|x1, x2, H)
× log2
∑
x′
2f(y|x1, x
′
2, H)
1S1
∑
x′
1
∑
x′
2f(y|x′1, x
′
2, H)
dy (4.40)
In the rest of this section, we refer to the mutual information in bits/dim.For this purpose, we define I ′r(H) as the mutual information in bits/dim as
I ′r(Hr) =1
L′rTdim
Ir(Hr) (4.41)
where L′r is the number of dimensions up to round r,(
∑rj=1 Lj = L′r
)
. We
can relate the generic throughput expression to the mutual information bydefining the probabilities in (4.19) as:
For a given channel realization hr and a particular value of SNR, themaximum rate of reliable communication supported by the channel at roundr is I ′r(hr) bits/s/Hz, which is a function of the random channel gain hr andis therefore random. If the transmitter encodes data at a rate Rr bits/s/Hz,then at round r, if the channel realization hr is such that I ′r(hr) < Rr,the transmission is called unsuccessful and this happens with probabilityPr(I ′r(hr) < Rr).
In this case, zero-outage is impossible since power control and channelstate feedback are not assumed [19]. However, we assume an outage con-straint at the end of the HARQ protocol. To model this constraint, we
50 Chapter 4 Mutual Information Analysis of Interference Networks
consider an IR-HARQ protocol with a maximum of Mmax rounds, and wesay that the constraint is met whenever the packet error probability afterMmax rounds is smaller than a predefined threshold Pout.
We look at the case of two transmissions rounds where for a given R1, wehave successful transmission in the first round if Pr(I ′1(H1) > R1), after thesecond round, outage corresponds to Pr(I ′2(H2) < R2). Let R1 = log2 B
L1Tdimbe
the rate at the first round, and R2 = log2 BNTTdim
the rate at the second round.Then, the overall throughput expression is:
R̄ =(
Pr(I ′1(H1)) > R1))
R1 +(
Pr(I ′1(H1) < R1, I′2(H2) > R2)
)
R2 (4.44)
where the outage probability is Pout = Pr(I ′1(H1) < R1, I′2(H2) < R2).
For the sake of obtaining long-term average throughput, we isolate thetarget channel h and average (4.44) over the channel distribution
Pr(
I ′1(H1) > R1
)
= EH Pr(
I ′1(H1) > R1|h)
(4.45)
To find the operating rates of the protocol, we start by choosing the ratein the second round as the rate that will satisfy the outage constraint, i.e. R2
that satisfies Pr(I ′2(H) < R2). The rate in the first round (R1), is chosen asthe one that maximizes the throughput expression in (4.44) while satisfyingthe given constraint. In this case, we also optimize the number of dimensionsused in each of the retransmission rounds.
Since there is no closed-form expression for the probability of outageof discrete signals, we notice that Pr(I(H2) < R2) represents the CDF ofthe mutual information evaluated at R2, i.e. FI(R2). With the help of theinversion formula in [62], we use the characteristic function of the mutualinformation ΦI(ω) to find the CDF as:
FI(R2) =1
2− 1
π
∫ ∞
0
ℑ{exp(−jωR2)ΦI(ω)}ω
dω (4.46)
The characteristic function is defined as ΦI(ω) = E [exp(jωI)], where E
denotes expectation. Since we assume Rayleigh fading, the expectation isover the channel squared magnitude probability density function (PDF),which is exponentially distributed:
ΦI(ω) =
∫ ∞
0exp(jωI(h)) exp(−h)dh (4.47)
Finally, we can use (4.46) and the outage constraint to solve the outageprobability expression for R2.
We start with the case of fixed rates across rounds when R1 = R2, andwe want to show the existence of one optimal rate for the protocols. Forthis purpose, we compute the throughput for the different cases by settingdifferent rates Rr at round r.
4.5 Practical Interference Networks Analysis 51
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Rate
No interference
1 round
2 rounds
3 rounds
SNR = 10 dB
(a) Reliable throughput, no interference
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 20
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
Rate
Interference
1 round
2 rounds
3 rounds
SNR=10 dB
(b) Reliable throughput, 1 interferer.
Figure 4.9: In (a) we have the reliable throughput for the HARQ protocolwith different number of rounds under Rayleigh fading, constant overrounds, and without interference, (b) shows the reliable throughput
different number of rounds and one dominant interferer. The rates acrossrounds are fixed and the operating SNR is 10 dB.
52 Chapter 4 Mutual Information Analysis of Interference Networks
In figure 4.9, we plot the reliable throughput for the HARQ protocolwithout interference and one strong interferer. We can identify that thereexists a value of the rate that gives the highest throughput or spectral effi-ciency. The latter means that there is an optimal value for the rate. If welook at the curves for one transmission round, we can observe an optimalrate of around 1.8 without interference and 1.3 with one interferer. We canalso see that with more than one transmission round, the optimal rate ishigher but after two rounds, it does not change. From these results, we cansee that if we let the maximum number of transmission rounds to increase,one could have the option to fix the coding rate to a value close to the maxi-mum rate and let the HARQ protocol work across the rounds until successfuldecoding. However, the latency in this case would be high. If we performrate adaptation, we can choose the optimal rate in each of the HARQ roundsand decode the information sooner, therefore minimizing the latency.
4.5.1 Rate Optimization (fixed across rounds)
We can now proceed to optimize the rate for different operating SNR points.This means that at every SNR point, we choose the rate that gives themaximum throughput. In this case, we do not consider an outage constraint.
In figure 4.10, we show the spectral efficiency and the probability ofoutage when we optimize the rate per transmission. The rate is the sameacross rounds. We compare rate optimization with a fixed rate operationfor a maximum number of HARQ rounds Mmax = 1, 2 and we can see that,for example, optimizing the rate with one HARQ round gives more or lessthe same gain as having an additional transmission round, but minimizingthe delay. If we allow two HARQ rounds with rate optimization, then thegain in throughput is even higher. In this case, the channel is constant andassumed unknown to the transmitter.
Figure 4.11 shows the spectral efficiency and the probability of outagewith one or two transmission rounds in a Rayleigh fading channel with QPSKmodulation. Across the transmission rounds, the rates are fixed. We considerthat there is one interferer present all the time, so the activity factor µ = 1.If we allow the activity factor to take other values in time than one, then wecan talk about the DL channel.
As a next step, we proceed to optimize the number of dimensions usedin each of the HARQ rounds.
4.5.2 Rate Optimization with an Outage Constraint
We first give an example on how to obtain the rates for the case of iidchannels across the HARQ rounds. According to the rate definition in (4.18),we define the rates R1 and R2 as R1 = log2 B
TdimL1and R2 = log2 B
Tdim(L1+L2)where
4.5 Practical Interference Networks Analysis 53
−10 −5 0 5 10 15 200
0.5
1
1.5
2
SNR [dB]
Th
rou
gh
pu
t
No interference
−10 −5 0 5 10 15 200
0.5
1
1.5
2
Ropt
2 rounds, Ropt
1 round, Ropt
2 rounds, no opt
1 round, no opt
Opt code rate
(a) Rate optimization, no interference.
−10 −5 0 5 10 15 2010
−2
10−1
100
SNR [dB]
Po
ut
No interference
1 round, no opt
2 rounds, no opt1 round, R
opt
2 rounds, Ropt
(b) Probability of outage, no interference
Figure 4.10: In (a) we show the rate optimization of the HARQ protocolfor a diferent number of rounds Mmax = 1, 2 in a Rayleigh fading channelwith QPSK modulation. The rates are fixed across rounds R1 = R2 = R.
Figure (b) shows the corresponding probability of outage.
54 Chapter 4 Mutual Information Analysis of Interference Networks
−10 −5 0 5 10 15 200
0.5
1
1.5
2
SNR [dB]
Th
rou
gh
pu
t
Interference
−10 −5 0 5 10 15 200
0.5
1
1.5
2
2 rounds, Ropt
1 round, Ropt
2 rounds, no opt
1 round, no opt
Opt code rate
Ropt
(a) Rate optimization, one interferer.
−10 −5 0 5 10 15 2010
−3
10−2
10−1
100
SNR [dB]
Po
ut
1 interferer
1 round, no opt
2 rounds, no opt
1 round, Ropt
2 round, Ropt
(b) Probability of outage, one interferer
Figure 4.11: In (a) we show the rate optimization of the HARQ protocolfor a diferent number of rounds Mmax = 1, 2 in a Rayleigh fading channelwith QPSK modulation. The rates are fixed across rounds R1 = R2 = R.There is one interferer all the time. Figure (b) shows the corresponding
probability of outage.
4.5 Practical Interference Networks Analysis 55
Lr is the number of dimensions used in round r. If we define ρ = L1L1+L2
(i.e.R2 = ρR1), then, the mutual information in bits/dim per round is given by:
I ′1(H1) =TdimL1I1(H1)
TdimL1= I1(H1) (4.48)
I ′2(H2) =Tdim × (L1I1(H1) + L2I2(H2))
Tdim(L1 + L2)= I1(H1) + I2(H2)
(
1− ρ
ρ
)
(4.49)
In this case, the characteristic function of the mutual information at thesecond round is given by:
φI2(ω) = E [exp(jωI ′2(H2))] (4.50)
= E
[
exp
(
jω
(
I1(H1) + I2(H2)
(
1− ρ
ρ
)))]
= E
[
exp (jωI1(H1)) exp
(
jωI2(H2)
(
1− ρ
ρ
))]
(4.51)
= E [exp (jωI1(H1))]E[
exp (jωI2(H2))( 1−ρ
ρ )]
(4.52)
= φI1(ω)φI2
(
ω
(
1− ρ
ρ
))
where going from (4.51) to (4.52) is possible because the channels are inde-pendent across rounds. Finally, we can use (4.46) and the outage constraintto find R2.
In figure 4.12, we show the results for the rate optimization with anoutage constraint of 10% and 1% when the channel remains constant acrossthe HARQ rounds. There is a clear advantage on optimizing the rate acrossthe HARQ rounds with a maximum gain of more than 10dB for the constantchannel case. The gain is higher when we have a more strict outage constraintof 1%. If we compare these results to those in section 4.5.1, without outageconstraint, we can see the significant gain introduced by optimizing the ratefor latency-constrained scenarios. It can also be noticed that the throughputwith rate optimization and an outage constraint of 1% is only lower bya small quantity as compared to the outage constraint of 10%. This lastobservation tells us that optimizing the rate can, indirectly, minimize theoutage constraint (achieving almost the same throughput for both outageconstraints).
In figure 4.13, we show the case for iid channels across transmissionrounds and the two outage constraints as in the previous case. To obtain themaximum throughput, we also investigate the case of removing the constrainton the modulation. To do this, we choose the rate in the first round accord-ing to the mutual information expression for Gaussian inputs, which is notbounded by a particular modulation order, and we choose the modulation
56 Chapter 4 Mutual Information Analysis of Interference Networks
−5 0 5 10 150
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR [dB]
2 rounds, constant channel, Pout
=0.1
Ropt
No optimization
7dB
(a) Rate optimization with an Outage constraint of 10%.
−5 0 5 10 150
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
SNR [dB]
2 rounds, constant channel, Pout
=0.01
Ropt
No optimization
> 10dB
(b) Rate optimization with an Outage constraint of 1%.
Figure 4.12: In (a) we show the rate optimization with an outageconstraint of 10%. The channel is constant across the HARQ rounds andthere is no interference. This is equivalent to a Non-Line-Of-Sight (NLOS)with slow fading channel. Figure (b) shows the corresponding curves for an
outage constraint of 1%
4.5 Practical Interference Networks Analysis 57
that allows us to achieve this rate. We define threshold values for changingmodulation between QPSK, 16-QAM and 64-QAM according to the maxi-mum rate achieved with each modulation for a particular SNR value. Weconfirm once more that the gain is higher for the 1% case (around 7dB forthe case of a more strict outage constraint of 1% while a gain of around2dB is observed fot the 10% case). When we change the modulation withrespect to the SNR, we observe a higher throughput in the high SNR region.This is caused by allowing the protocol to use higher modulation orders.In the following section, we show results from bringing interference into thescenario.
Interference case
We consider one dominant interferer and we assume a constant channel onthe desired user. We analyze the case for the UL and DL differently. Wefirst consider the DL of a femtocell with one interferer, and we consider anactivity factor for the interferer which means that it is active only a portionof the time. We consider the channel of the femtocell user as AWGN sinceit is likely to experience a good channel. For the interference, we considerthe channel as Rayleigh. Figure 4.14 shows the results for an activity factorof 50%, i.e. interference is present only half of the time. We fix the outageconstraint at 1% and we plot the throughput for the user of interest. Infigure 4.15, we show the DL case for an outage constraint of 10%. We observethe same behavior as the case without interference, with a higher gain fora more strict outage constraint (2dB for the 10% case against 10dB for the1% case). For the UL, since the interference is coming from other users,it changes in time. Therefore, we consider the interference constant and iidacross the HARQ rounds. Figure 4.16 shows the results for rate optimizationin the 10% outage constraint case.
In this chapter, we have demonstrated the benefits of adapting the rateand physical dimensions across transmission rounds of HARQ protocols. Weobtained a throughput higher than the ergodic capacity in the case of zero-outage throughput and we have showed that having an upper layer ARQin case of a residual outage probability results in a lower throughput. Inpractical scenarios without power control and channel state information, itis not possible to get zero-outage throughput. However, we benefit fromthe dynamic resource allocation and by imposing a constraint on the outageprobability, we can improve the throughput by varying the latency of theprotocol. By motivating the use of inter-round resource allocation, we de-rived distributed resource allocation policies that are applicable for the ULand DL channels. We considered interference that can be bursty due to thecharacteristics of HetNets deployments. Rather than performing extensivesimulations, we derived analytical expressions, based on mutual information
58 Chapter 4 Mutual Information Analysis of Interference Networks
−5 0 5 10 150
0.5
1
1.5
2
2.5
3
3.5
4
SNR [dB]
2 rounds, iid channels, Pout
=0.1
Ropt
, R1Gaussian
Ropt
, QPSK
opt. R,R/2 QPSK
QPSK
16−QAM
64−QAM
2dB
(a) Rate optimization with an Outage constraint of 10%.
−5 0 5 10 150
0.5
1
1.5
2
2.5
3
3.5
4
SNR [dB]
2 rounds, iid channel Pout
=0.01
Ropt
, R1Gaussian
Ropt
, QPSK
opt R, R/2 QPSK
QPSK
16−QAM
64−QAM
7dB
(b) Rate optimization with an Outage constraint of 1%.
Figure 4.13: In (a) we show the rate optimization with an outageconstraint of 10%. The channel is iid across the HARQ rounds and there isno interference. In this case, it is equivalent to a NLOS channel with fast
fading or frequency hopping. Figure (b) shows the corresponding curves foran outage constraint of 1%
4.5 Practical Interference Networks Analysis 59
−5 0 5 10 150
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR [dB]
Downlink, Pout
=0.01, activity factor 50%
Ropt
, QPSK
QPSK, no optimization
10dB
Figure 4.14: Rate optimization for Rayleigh fading on the downlinkchannel. There is one dominant interferer with an activity factor of 50%.
The outage constraint is 1%
−5 0 5 10 150
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR [dB]
Downlink, Pout
=0.1, activity factor 50%
Ropt
, QPSK
QPSK, no optimization
2dB
Figure 4.15: Rate optimization for Rayleigh fading on the downlinkchannel. There is one dominant interferer with an activity factor of 50%.
The outage constraint is 10%
60 Chapter 4 Mutual Information Analysis of Interference Networks
−5 0 5 10 150
0.2
0.4
0.6
0.8
1
1.2
1.4
SNR [dB]
Uplink, Pout
=0.1
opt R1 , R
2
no optimization
2dB
Figure 4.16: Rate optimization for Rayleigh fading on the uplink channel.There is one dominant iid interferer. The outage constraint is 10%
modeling, that capture the throughput performance of the network with orwithout a latency constraint.
4.6 Methodology for Resource Allocation in Prac-
tical Interference Scenarios
To illustrate some of the applications of the analytical framework presentedin Chapter 4, we provide in the following sections several examples in morepractical scenarios. First, we present a Manhattan-like topology for the caseof small-cells inside a block of apartments. Then, we present a cross-layerinterference case, with a macrocell overlaid by a single femtocell and we takea look at the DL channel of the macrocell user. Finally, we explain theprocedure to perform PHY abstraction for interference scenarios. It shouldbe noted that all the results in this section are using the rate optimizationprocedures described in section 4.4.2.
4.6.1 Manhattan-topology for Small-cells
We can model small-cell networks with non-constant interference by usingthe activity factors. A block of apartments is modeled as a rectangular gridwith small-cells positioned at the center of each apartment. If we assumethat the small-cells are active with a certain probability, then the interferenceprocess is no longer ergodic and interference is randomized in time. Westudy the performance of our rate adaptation policy in this apartment blockscenario. For this purpose, we look at the DL of the small-cells Manhattan-
4.6 Methodology for Resource Allocation in Practical Interference Scenarios61
Figure 4.17: Manhattan-like topology with the user of interest at the edgeof the apartment and the rest of the interfering small-cells placed at the
middle of the surrounding apartments
like topology in figure 4.17 with the user of interest located inside the uppercentral apartment. To model indoor scenarios, we consider the path lossexponent in (4.21) to be α = 3 [41]. The size of the apartments is 10×10m2
and we measure, at every point, the distance to the neighboring interferersand scale the noise accordingly. We assume that all nodes transmit at equalpower and the interference is known at every slot, i.e. we assume that theconstellation and the channel can be estimated. This can be achieved, forinstance, through a loose-cooperation policy where the base stations agreeon the bandwidth and constellation they use.
Figure 4.18 shows the throughput of the small-cell user under Rayleighfading, located along a 45 ◦ line from the center towards the corner of theapartment. The neighboring small-cells are located at the center of everyapartment. We take into account one dominant interferer. In the case thereare three interferers located at the same distance, we consider the maximumof three Rayleigh random variables and we apply a Gaussian approximationto the rest of the interferers by scaling the noise with the correspondingpower. We consider QPSK 16-QAM, and 64-QAM modulations and an ac-tivity factor of 50%. We assume that, at the corner, the SNR for the useris equal to the strength of the interferer and is 10dB and we consideredan attenuation loss (penetration of walls/floors) of 10dB. As expected, thethroughput decreases proportionally to the distance from the small-cell. Indistances close to the small-cell, the throughput with 64-QAM is substan-tially higher than when using 16-QAM or QPSK. After 5.5 m, they all achievethe same throughput.
62 Chapter 4 Mutual Information Analysis of Interference Networks
1 2 3 4 5 6 7 80
1
2
3
4
5
6
distance [m]
Throughput, Pout
=0.01
64QAM
16QAM
QPSK
Figure 4.18: Downlink of a small-cell user. Interference is coming from theneighboring small-cells with an activity factor µ = 50%. The modulation is
QPSK, 16-QAM, or 64-QAM and the SNR is equal to the interferencestrength and is 10dB. We consider the strongest interferer and we make aGaussian approximation for the rest. The rate is adapted across rounds for
all cases.
4.6 Methodology for Resource Allocation in Practical Interference Scenarios63
Figure 4.19: Downlink of a macrocell with a femtocell interfering.
4.6.2 Macro/Small-cell Scenario
In this case, we look at a scenario where the cellular network is overlaidby a femtocell network. We model a scenario where a single macrocell isoverlaid by one femtocell and we look at the DL of the macrocell user whenthe interference is coming from the interfering femtocell (see figure 4.19).
−5 0 5 10 150
0.2
0.4
0.6
0.8
1
1.2
1.4
SNR [dB]
Downlink of a macro user, interference from femto
µ = 30%
µ = 50%
µ = 70% µ = 50%
QPSK macro user
16−QAM femtocell
QPSK−QPSK
QSPK−16−QAM
Figure 4.20: Downlink of a QPSK macro cell user under Rayleigh fading.Interference is coming from a 16-QAM femtocell active only part of thetime. Curves are shown for different activity factors 30%, 50% and 70%.
We model a femtocell that is active only with a certain probability specifiedby the activity factor. We consider an activity factor of 30%, 50% and 70%.The macro cell user is considered to be active all the time. Figure 4.20shows the results when the modulation for the macro cell user is QPSK andfor the interfering femtocell, we consider 16-QAM, since a femtocell user will
64 Chapter 4 Mutual Information Analysis of Interference Networks
experience a better channel. For an activity factor of 50%, we make thecomparison of a femtocell using either QPSK or 16-QAM and we observethat for low to moderate SNR, the achieved throughput is very close. Thehigher constellation of the interference results in a lower throughput onlyafter 10 dB. We have to remember that we assumed the strength of theinterference equal to the user of interest.
4.6.3 Physical layer Abstraction Models
Simulating an entire physical layer can be extremely time consuming andin some cases computationally unfeasible. By using analytical expressionsthat model the performance of interference networks, we can abstract thephysical layer. PHY abstraction methods allow us to accurately predict thecoded Block Error Rate (BLER) for any given channel realization.
The BLER can be obtained through mapping mechanisms from instanta-neous channel state measures such as SINR. When the mapping mechanismis a function of the mutual information capacity, it is called Mutual Informa-tion Effective SINR Mapping (MIESM) and it makes use of stored look-uptables. Performing PHY abstraction is out of the scope of this work, how-ever, we explain the procedure that has to be followed in order to do theabstraction.
The first step to perform a PHY abstraction is to compute the SINRgiven a set of parameters such as power, interference, fading, etc. This SINRvalue is then used to compute the corresponding mutual information. Themutual information will be different at every round of the HARQ protocolsince we use a different number of dimensions per round and the interferenceconditions may vary (due to the burstiness of the traffic). Finally, we haveto map the mutual information to the corresponding BLER and store it ina look-up table. The effective SINR mapping for MIESM is given by:
SINReff = β1f−1
1
P
P∑
p=1
f
(
SINRp
β2
)
(4.53)
where the f function is a mapping function to an information metric, i.e.the mutual information expression for discrete constellations, β1,2 are cali-bration factors that have to be adjusted according to the channel model andmodulation schemes used. The look-up table is then constructed mappingthe mutual information values to the corresponding BLER (or outage prob-ability). We have derived the necessary expressions to compute the BLERfor a range of mutual information values.
4.6 Methodology for Resource Allocation in Practical Interference Scenarios65
In the next chapter, we investigate the performance of our resource allo-cation strategies on practical LTE MODEMs in an interference environmentunder the constraints of LTE coded-modulation.
66 Chapter 4 Mutual Information Analysis of Interference Networks
Chapter 5
Practical Scheduler Design for
LTE Base Stations
A key feature of LTE is the adoption of advanced radio resource managementprocedures in order to increase the system performance. The efficient use ofradio resources is essential to meet the system performance targets [21]. Inthis chapter, we study the performance of our resource allocation policies forthe scheduling of HARQ transmissions under the constraints of LTE coded-modulation.
The HARQ and rate adaptation algorithm is implemented and managedby the scheduler at the MAC layer. The scheduling mechanism plays afundamental role because it is responsible for choosing, with fine time andfrequency granularity, how to distribute the radio resources among differentusers, taking into account the current channel conditions and QoS require-ments. The scheduling algorithm is a base station-implementation issue andit is not specified in the standard. The standard only gives channel-qualitymeasurements/reports and specific procedures for dynamic resource alloca-tion [28] needed for the terminal’s implementation (and consequently thebase station as well). This allows to have vendor-specific algorithms whichcan be optimized for specific scenarios.
The design of effective resource allocation mechanisms becomes crucialfor an efficient network operation. Low complexity and scalability are funda-mental requirements for finding the best allocation decision. Complex andnon-linear optimization problems or exhaustive research over all the pos-sible combinations would be too expensive in terms of computational costand time [45]. An algorithm should be easily implemented and, above all,should require very low computational cost. Many theoretical solutions can
67
68 Chapter 5 Practical Scheduler Design for LTE Base Stations
be found in the literature, but when studied closely, they cannot be deployedin real systems given the difficulty to be implemented in real devices and thehigh computational cost required. Robust strategies should guarantee theability to work in very different scenarios. It should not require strong pa-rameter settings, or it should at least dynamically adapt such parameters toenvironmental changes [21].
5.1 Handling Interference in LTE Networks with
HARQ and AMC
Traditionally, HARQ has been used as a way to recover from errors oc-curring during the transmission of information, decreasing the probabilityof unsuccessful decoding. HARQ can be combined with Adaptive Modula-tion and Coding (AMC). When implementing AMC with HARQ, the MCSis adapted between each retransmission. The coding rate can be adaptedacross transmissions to further improve the performance. The code rate canbe fine-tuned by puncturing, generating different redundancy versions tomatch the number of coded bits to the channel. The code rate, rate match-ing, and the number of resources allocated for one transmission determinethe transport-block size (TBS) [61]. In essence, rate adaptation adapts theMCS to the current channel conditions which translates to the data rateor error probability of each link. The MCS along with signaling overhead(reference signals, control channel resource elements), in terms of spectralefficiency, represents the number of information bits per modulation symbol.The use of rate adaptation provides manufacturers an incentive to implementmore advanced receivers since it will result in higher end-user data rates thanstandard receivers [27].
Based on the available CQI, it is the scheduler at the eNB that canaddress the different quality and service requirements of all the associatedUEs in a spectrally efficient way.
LTE also supports the adaptation of the so-called TBS and the amountof PRBs used per transmission. A transport-block is the name given toa block of data at the MAC layer [9] and corresponds to a codeword, ormore precisely a set of codewords, at the physical layer. For LTE, HARQ issupported for the Physical Uplink Shared Channels (PUSCH) in both UL andDL (PDSCH), and separate control channels with configurable levels of errorprobabilities on the uplink are used to send the associated acknowledgmentfeedback. HARQ can be classified as either synchronous or asynchronous andthe retransmissions can be adaptive or non-adaptive [61]. In a synchronoussystem, retransmissions occur at a predefined time. In an asynchronoussystem, the retransmissions can occur at any time (and must be signaled).With adaptive HARQ, the MCS and other transmission attributes can bechanged after each round. In a non-adaptive context, transmission attributes
5.2 OpenAirInterface Implementation 69
are fixed or pre-defined. In LTE, HARQ is asynchronous and adaptive in theDL and synchronous in the UL. Retransmissions can be adaptive or not inthe UL [61]. Feedback for HARQ in LTE comprises a simple ACK/NACKsignal.
An important factor for the eNB scheduling algorithm is the accuracy ofthe available CQI for the active UEs in the cell. In the DL, CQI is reportedback by the UE. For the UL, the eNB can use Sounding Reference Signals(SRS) or other signals transmitted by the UEs to estimate CQI [61]. Onthe UL, two channels are used to send CQI reports. Namely the PhysicalUplink Control Channel (PUCCH) and the PUSCH. Reporting can be peri-odic or aperiodic. Periodic reports are normally transmitted on the PUCCH.However, the eNB can request the UE to send aperiodic CQI reports on thePUSCH because it is more appropriate to transmit large and detailed re-ports [38]. The key issue with respect to resolving channel quality is theability to obtain accurate information in the presence of a large number ofconnected UEs (even idle) and the presence of sporadic interference (primar-ily on the uplink). The eNB can adjust the periodicity and granularity ofthe CQI feedback (down to 2 ms periodicity for both CQI and SRS in Rel-10LTE [7]) allowing it to trade-off between the amount of overhead and theaccuracy of the channel information. Of course, when a long delay occurswith respect to the scheduling time, the performance can be significantlyaffected. Short feedback periodicity is difficult to achieve in heavily loadedcells.
In summary, LTE offers a lot of flexibility in terms of resource allocationand, in particular, resource allocation algorithms can be tailored for a par-ticular class of traffic with specific requirements. Nevertheless, work is stillneeded to exploit this flexibility efficiently for key emerging applications. Tothis end, we consider the dynamic resource allocation schemes for IR-HARQpresented in Chapter 4 under the constraints of LTE coded-modulation. Ourgoal is to analyze the feasibility of such techniques in real LTE MODEMsand to measure the performance and possible limitations with respect to theimplementation in real systems. Furthermore, we want to measure how closewe can get to theoretical predictions of the performance and distribution ofthe physical dimensions across HARQ rounds.
5.2 OpenAirInterface Implementation
We first present the OAI SDR platform used to test the performance andcompliance of our resource allocation strategies in LTE. OAI is an open-source hardware/software development platform and open-forum for innova-tion in the area of digital radio communications [3]. It was created by theMobile Communications Department at EURECOM based on its experiencein publicly-funded R&D carried out in the context of collaborative research
70 Chapter 5 Practical Scheduler Design for LTE Base Stations
projects [15]. One of the most significant benefits from a full SDR platformis that the same code can be used for emulation as in a real implementation,providing a smooth transition from the simulations to the real tests.
In the following sections, we give the detailed procedure for the resourceallocation algorithm implementation on the downlink channel.
5.2.1 Physical Layer and Resource Allocation
In LTE, data is encapsulated in the so-called Packet Data Units (PDUs) atthe MAC layer and forwarded to the PHY for transmission over the air. Thephysical layer provides services to the MAC layer in the form of transportchannels and is responsible for the coding, modulation, and TBS determina-tion and assignment of the physical resources for transmission of information.LTE defines a number of downlink physical channels to carry informationblocks received from the MAC and higher layers. These channels are cat-egorized as transport or control channels. In the DL, the PDSCH is themain physical channel allocated to users on a dynamic and opportunisticbasis. It carries all user and control-plane information and it is scheduledusing the Downlink Control Information (DCI) on the Physical DownlinkControl Channel (PDCCH). The UE decodes the PDCCH every sub-frameto know which PDSCH to decode and where to find them in time or fre-quency. There are different formats for the DCI, but we use DCI format 1defined for the scheduling of one PDSCH codeword (Transmission modes 1,2 or 7). The DCI includes information about the resource allocation type,RBs assignment, MCS, HARQ information and redundancy version, and thepower control command for the PUCCH. The PDSCH carries data in TBs.They are passed from the MAC layer once per TTI which in LTE is 1 ms.The physical resources are assigned on a basis of two resource blocks for oneTTI. This is referred to by “pair of resource blocks” which is the minimumof resources that can be allocated.
In general, a compromise between flexibility and signaling overhead hasto be met when sending the resource allocation assignments in terms of aset of RBs. The simplest solution would be to send a bitmap for each UE inwhich each bit indicates a particular RB, but this would be costly for largebandwidths (> 5 MHz). The resource allocation in DCI is described by twoparts: the header (type of allocation : RA type 0,1,2), and the resource blockassignment itself. We give the details of the resource allocation type that weadopted which is part of the LTE standard Rel-8.
Resource Allocation Type 0
As mentioned before, the bandwidth is divided into groups of RBs and abitmap indicates the actual resource block assignment, i.e. the ResourceBlock Groups (RBGs) that are allocated to the scheduled UE, where RBG
5.3 Application of the Scheduling Policies in LTE 71
is a set of consecutive PRBs. The RBG size (P) is a function of the systembandwidth (see Figure 5.1 [7]). The total number of RBGs (NRBG) for
a downlink system bandwidth of NDLRB PRBs is given by NRBG =
⌈
NDLRBP
⌉
where⌊
NDLRBP
⌋
of the RBGs are of size P and if NRBG mod P> 0 then one
of the RBGs is of size NRBG − P ∗⌊
NDLRBP
⌋
.
Table 5.1: RBG Size vs Downlink System Bandwidth
System Bandwidth RBG size(NDL
RB) (P)
≤ 10 111-26 227-63 364-110 4
Modulation and TBS Determination
Modulation and coding formats are limited with respect to achieving allspectral efficiencies (which is possible with LTE rate matching), but thegranularity is quite fine. There are 5 bits that provide a table entry (IMCS)giving the modulation order (Qm) : QPSK (2), 16-QAM (4), 64-QAM (6)and TBS indication (ITBS). This information is combined with the numberof allocated RBs to yield the transport block size (table lookup). The UE inthe physical layer delivers ITBS along with HARQ process ID (from PDCCH)and New Data Indicator (NDI) bit to the MAC layer in parallel to decodingPDSCH/DLSCH.
The selection of the TBS, MCS and antenna mapping, together with thelogical-channel multiplexing for downlink transmissions are controlled by theeNB. In the case of a retransmission, the TBS is, by definition, unchanged,but the same TBS should ideally appear for several different resource-blockallocations as this allows the number of resource blocks to be changed be-tween retransmission attempts, providing increased scheduling flexibility [7].
5.3 Application of the Scheduling Policies in LTE
The policies considered in the previous chapters require an incremental-redundancy coded-modulation system with the possibility of adjusting thecode rate (physical resources) via puncturing and repetition (rate matching).This is possible on the LTE Rel-8/10 DL and UL where the allocation can beadapted across transmission rounds, the granularity of which depends on thetransport block size. The only difference between the UL and DL allocation
72 Chapter 5 Practical Scheduler Design for LTE Base Stations
is that the modulation order must remain fixed on the UL for each round.It is likely that this restriction is insignificant with respect to performance.The slight penalty for inter-round rate adaptation is the requirement to senda new DCI packet with updated resources rather than an automatic retrans-mission. Let us now explain with an example the rate-adaptation policy fortwo-round transmission in an LTE context. For simplicity, we consider thecase with Gaussian signals and without interference in Chapter 4.3.
For illustration purposes, consider a DL transmission in transmissionmode 1 (SISO transmission). The policy must choose the initial spectralefficiency, MCS1 corresponding to the first transmission round and N1 andN2 corresponding to the number of allocated physical resource blocks. Anadditional constraint (on the downlink) is that these must be multiples ofan integer P which is the allocation granularity dictated by the transmissionbandwidth (for a 10MHz carrier, P = 3 [7].) Figure 5.1 shows the resultof an optimized two-round protocol, specifically the rate R1 which must betranslated to MCS1 and ρ which provides the ratio of the physical dimensionsused in the two rounds. This can now be transformed as a function of thetarget transport block size. Now consider an average SNR of 7 dB, thiscorresponds to R1 of 2 bits/dimension and ρ = 0.07. Assume an allocationin the first round using 3 resource blocks (minimal allocation for a 10 MHzcarrier). In a PDSCH-only normal prefix subframe with one PDCCH symbol,the total number of resource elements is N1 = 450 (13 PDSCH symbols, 3with 10 PDSCH resource elements per resource block and 10 with 12). Thetarget transport block size would therefore be around 900 bits, so the closesttransport block size for 3 resource blocks is 904 bits ( [7, Table 7.1.7.2.1-1])with MCS1 = 16 (16QAM). The second-round dimensions would ideally beN2 = N1(
1−ρρ
) = 5978 yielding 39.9 (39 or 42) resource blocks.
The previous example exposes the difficulty of applying such latency-constrained policies for large transport block sizes, at least under the con-straints of the current LTE specifications. For instance, if we were to use thesame operating point (7 dB average SNR) with twice the number of dimen-sions in the first round, corresponding to a transport block size of 1800 bits,the number of required resource blocks in the second round jumps to 78 or 81which is only realizable on a 20 MHz carrier. This can be easily overcome byallowing a third transmission round, i.e. by increasing the acceptable latencyof the transmission. Another important consideration regarding applicationon the UL is that the nature of the resource allocation policy is fundamentallyrelated to power control, since we are assuming a constant transmit energyper channel dimension. This is also the adopted policy in LTE (assumingpower adjustments are not made during retransmission rounds). Basically,low power is used during the first transmission and significantly more poweris used in the second transmission if required. In the numerical example
5.3 Application of the Scheduling Policies in LTE 73
0 5 10 15 20 25 300
1
2
3
4
5
6
7
8
9
SNR [dB]
R1
(a)
0 5 10 15 20 25 300
0.1
0.2
0.3
0.4
0.5
0.6
0.7
SNR [dB]
ρ
(b)
Figure 5.1: In (a), we consider the scenario without CQI (uncorrelatedchannels), and we plot the rate in the first round (R1) for different valuesof SNR. We fix the probability of outage after the second round to 1%.
In (b), for the scenario without CQI (uncorrelated channels), we plot the ρparameter against different values of SNR. We fix the probability of outage
after the second round to 1%. ρ determines the rate used in the secondround according to equation [4.6]
74 Chapter 5 Practical Scheduler Design for LTE Base Stations
above the power boost is 10 log10
(
1−ρρ
)
= 11.2dB. This, of course, requires
that the UE has signaled sufficient power headroom [7] for the eNB to allowthis allocation. This clearly shows that, on the UL, latency can be controlledthrough a combination of rate adaptation, HARQ and power control. Aninstance of this appearing in the literature in the case of MIMO transmissionwith HARQ can be found in [29], although in that example the number ofdimensions across tranmsission rounds remained fixed and the energy perchannel dimension increased across rounds.
5.4 Performance Analysis of the Scheduler
A scheduling algorithm responds to a pre-formulated capacity-related metricwhich is optimized across all possible resource allocation solutions satisfyinga set of predetermined requirements such as QoS, spectral efficiency, or la-tency. To compute the spectral efficiency, we have to consider the TBS, themodulation order and the outage probability:
Speceff =Qm,1
G1TBS
(
1−Pout,1
)
+
(
Qm,1
G1+
Qm,2
G2
)
TBS(
Pout,1(1−Pout,2))
(5.1)where Gi is the number of coded bits per codeword, and Qm,i the modulationorder used at the ith round, previously defined in section 5.2.1. Pout,1 is theoutage probability after the first round, and Pout,2 is the outage probabilityafter the second round.
As an example of the optimization of (5.1), we fix the TBS for all pos-sible allocations to around 1000 bits and from the table 7.1.7.2.1-1 in [7],we get the TBS index depending on the number of PRBs used in the firstround and the TBS closest to 1000. Finally, table 7.1.7.1-1 (see [7]) gives usthe corresponding MCS index and modulation order. Table 5.2 shows thedifferent allocations used to generate the results in figures 5.2 and 5.3, where(
NDLRB
)
1represents the number of PRBs allocated in the first round. We
use this table to test different allocations and investigate the performanceof the LTE codes. By using the DCI, we can signal the new informationwith respect to the number of PRBs in the consecutive rounds of the HARQprotocol. The MCS can also be adapted, but the TBS remains fixed.
5.4 Performance Analysis of the Scheduler 75
Table 5.2: Resource Allocation vs Modulation and Transport Block Sizes
Figure 5.3: Probability of Outage after two HARQ rounds. There is oneinterferer of the same strength as the user of interest with a 50%probability of being active. The channel is AWGN for both users.
We can see in figures 5.2 and 5.3 the performance of the LTE constrainedcoded-modulation without interference and with one dominant interfererwhich is active 50% of the time. Theoretically, all the codes should behave inthe same way, meaning that they should have the same outage probability.However, the results show that for the smallest allocations and therefore thehighest modulations (MCS values higher than 11) there is a difference withrespect to the lower MCS values. If we consider the modulation used bythese MCS values, we can see that MCS = 11, 12, 14 use 16-QAM and MCS23 uses 64-QAM modulation. In the case of interference shown in figure 5.3,we show also the performance of the codes when forcing the retransmissionsof these MCSs to use QPSK (MCS 29) or 16-QAM (MCS 30). As we see,in general, the best result is achieved by using QPSK in the second round(except for MCS 23 where is better to use 16-QAM). In our results, to re-duce the gap for different MCSs, we force the retransmissions to use QPSKby setting the MCS value to 29 as indicated in the 3GPP standard (MCSvalues 29, 30, 31 are reserved in the LTE standard to adapt the modulationorder across rounds). This has no impact on spectral efficiency but changesthe behavior of the rate-matching algorithm by allowing the code to operatewith less repetition which may be more efficient.
5.4 Performance Analysis of the Scheduler 77
In the following sections, we show the results for the optimization of thenumber of resources in the second round (the total number of PRBs used inboth rounds can be less than NDL
RB . To optimize this number, we follow thenext procedure:
1. fix the TBS (eg. 1000 bits)
2. run over the different MCS and SNR values
3. compute the outage in the first round
4. run over all combinations of PRBs in the second round until an outageof 1% is satisfied (we fixQm = 2 for retransmissions)
5. choose the corresponding rate at the second round R2 and computethe spectral efficiency over the rounds
In the following sections, we give the results of this procedure in the casewithout interference and one dominant interferer, where we also take intoaccount the activity factors which model bursty interference.
The combination of the fixed-rate turbo-code with dynamic resource allo-cation and the rate-matching permutation amounts to doing rateless codingsuch as those proposed in [30] and [58] where they consider an unboundednumber of transmission rounds. Our dynamic adaptation works over a smallnumber of transmission rounds (2-4) and can actually be applied to suchcoding schemes. The latter would require appropriate link-layer HARQ pro-tocols (e.g. [40]) adapted to such dynamic rateless coders.
It should be noted that all simulations in this chapter assume imperfectchannel estimation. If we would like to truly compare to rateless codes, weshould assume perfect channel estimation or channel state information atthe receiver (CSIR).
5.4.1 Results without interference
In this section, we present the results for the resource allocation strategieswith different channel models. We simulated 10000 frames for each of theSNR points and MCS values using the OAI unitary LTE PHY link simula-tor and we give the results in terms of the spectral efficiency. For referencepurposes, we also give the probability of outage after the first round Pout,1.The outage probability after the second round Pout,2 is fixed to 1% for allsimulations.
78 Chapter 5 Practical Scheduler Design for LTE Base Stations
AWGN channel
First, we present the case with AWGN channel and no interference. Weoptimize the allocation of PRBs in the second HARQ round by followingthe procedure described in section 5.4. We save the number of erroneoustransmissions after each round to compute the probability of outage and weuse it together with the number of coded bits per codeword to compute thespectral efficiency as in (5.1).
−6 −4 −2 0 2 4 6 810
−2
10−1
100
SNR [dB]
Pout1 (AWGN, no interf.)
mcs2
mcs3
mcs4
mcs5
mcs6
mcs7
mcs8
mcs9
mcs11
mcs12
Figure 5.4: Probability of Outage after the first round of the HARQprotocol. The channel is AWGN and there is no interference.
Figure 5.4 shows the results for the probability of outage after the firstHARQ round. In figure 5.5, the left axis and solid lines show the spectralefficiency depending on the resource allocation chosen. The right axis anddashed lines show the corresponding (optimized) number of PRBs used in thesecond round. We show the MCS used in each case (for the retransmissionsthe MCS is fixed to force them to use QPSK), the TBS remains fixed acrossthe HARQ rounds. We also show the spectral efficiency values specified inthe LTE standard for terminal feedback signaling (see Table 7.2.3-1 in [7]). If
5.4 Performance Analysis of the Scheduler 79
we look at the lines showing these values, we can see the corresponding MCSthat would be used in each case. For example, if we consider the spectralefficiency of 0.3770, we would be using MCS 3 which gives the highest achiev-able spectral efficiency. If we consider the next spectral efficiency 0.6016, wewould use MCS 6. However, for the large number of intermediate values thatthe protocol can use, one could change the allocation and/or MCS depend-ing on the SNR to maximize the spectral efficiency. This could be done bythe base station scheduler based on estimates of the first and second rounderror probabilities. The latter could be achieved based on the statistics ofreceived ACK/NACK signals. By doing so, one has more liberty to adaptparameters such as PRB allocation or MCS and modulation across roundsto those values that result in a higher spectral efficiency given the latencyconstraint.
The desired latency of the protocol translates as the probability of outagefor the different operating SNR points. For our latency constraint of 10−2,we can look at different SNR values and see what the best allocation is. Ifwe consider an SNR= 0dB, the corresponding MCS is 3 and the allocationcorresponds to 18 PRBs in the first round and 3 in the second round. Thiscorresponds to a ratio of dimensions of ρ = 0.8571. If we now consider anSNR= 5dB, the MCS becomes 9 with 6 PRBs in the first round and 2 in thesecond round (ρ = 0.75). As the SNR gets higher, we use less resources inthe second round. Interestingly, if we look at the outage probability after thefirst round, we see that the second case of MCS 9, has a higher probabilityof outage when compared to the case with MCS 3, which means that inthis case, almost all the data transmission occurs at the second round of theHARQ protocol.
Rayleigh Channel
We investigate the effect of the distribution of dimensions across rounds (re-lated to section 4.3.2, Chapter 4) and we look at the results when the channelis Rayleigh distributed giving a worse channel quality than the AWGN case.This is equivalent to the outdated CQI case with uncorrelated channels pre-sented in section 4.3.2 of Chapter 4.
Figure 5.6 show the probability of outage after the first round of theHARQ protocol. Figure 5.7 shows the spectral efficiency for the differentoptimized allocations and those specified in the LTE standard. When weconsider an SNR of 2dB, the best combination of MCS and PRBs allocationis MCS 7 with 8 PRBs in the first round and 25 in the second round, whichgives a ratio of dimensions ρ = 0.2424. However, if we look at high SNR(15dB), the best MCS becomes 16 with 3 PRBs in the first round and 6 inthe second round, with ρ = 0.3333.
80
Chapte
r5
Pra
ctic
alSch
edule
rD
esig
nfo
rLT
EB
ase
Sta
tions
−2 −1 0 1 2 3 4 5 6 7 80.25
0.35
0.45
0.55
0.65
0.75
0.85
0.95
1.05
1.15
1.251.3
SNR [dB]
Spectr
al eff
1
3
5
7
9
11
13
15
17
19
21
23
25
NP
RB
2
mcs2, NPRB1=23
mcs3, NPRB1=18
mcs4, NPRB1=14
mcs5, NPRB1=12
mcs6, NPRB1=10
mcs7, NPRB1=8
mcs8, NPRB1=7
mcs9, NPRB1=6
mcs11, NPRB1=6
mcs12, NPRB1=5
0.8770
0.6016
1.1758
0.3770
Figure 5.5: Spectral efficiency for MCS= {2, 3, 4, . . . , 9}. There is no interference.
5.4 Performance Analysis of the Scheduler 81
−5 0 5 10 15 2010
−2
10−1
100
SNR [dB]
Pout1 (Rayleigh, no interf)
mcs3
mcs4
mcs5
mcs7
mcs9
mcs13
mcs15
mcs16
mcs18
Figure 5.6: Probability of Outage after the first round of the HARQprotocol under Rayleigh fading. There is no interference.
82
Chapte
r5
Pra
ctic
alSch
edule
rD
esig
nfo
rLT
EB
ase
Sta
tions
−5 0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2
SNR [dB]
Spectral eff (Rayleigh, no interf)
mcs3
mcs4
mcs5
mcs7
mcs9
mcs13
mcs15
mcs16
mcs18
1.1758
1.9141
1.4766
0.8770
0.3770
0.6016
0.2344
0.1523
Figure 5.7: Spectral efficiency for different resource allocations under Rayleigh fading. There is no interference.
5.4 Performance Analysis of the Scheduler 83
2 4 6 8 10 12 14 16 180
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNR [dB]
ratio of dimensions (ρ)
ρ simulation
ρ Gaussian
Rayleigh channelno interference
unconstrained modulation (PERFECT CE)
LTE coded−modulation(IMPERFECT CE)
Figure 5.8: Ratio of dimensions.
We can compare these values with the theoretical example in figure 5.1(b).For this purpose, in figure 5.8, we plot the ratio of dimensions between firstand second round and we also plot the results from figure 5.1(b). The differ-ence in the curves comes from the fact that in the theoretical example, theresults were obtained from the Gaussian expressions for mutual informationwhich represent an unconstrained modulation and in the case of the simu-lation results, we study the LTE codes with imperfect channel estimation.However, we see that the results follow the same trend in both cases with ρbecoming higher in the high SNR region. Also, from figure 5.6, we observethat as the SNR becomes higher, the outage probability after the first roundbecomes higher. The latter is important if we consider that commercial sys-tems are designed to operate in the order of 10−1, but allowing a higheroutage probability results in a higher spectral efficiency. We can relate ourresults to those from the rateless coding schemes with AWGN channels. Inthis case, more dimensions are used in the first round than in the secondsince it is possible to overcome errors due to finite block-length (inducing agap from the Shannon capacity).
84 Chapter 5 Practical Scheduler Design for LTE Base Stations
Figure 5.9: Scenario.
5.4.2 Results with one interferer
In this section, we present the results for the interference case in figure 5.9.We consider one dominant interferer with the same power strength as the userof interest and an activity factor of 50% which means that it is active only halfof the time. We performed simulations for different channel models for boththe user of interest and the interference. We followed the same procedureto optimize the resource allocation as for the case with no interference. Wegenerate data signals for both the user and the interferer and we modelexplicitly the interference by adding both signals. At the receiver we donot perform any type of interference cancellation. However, we observe thatusing the best resource allocation decreases, to a certain extent, the impactof the interference.
AWGN / AWGN
For the simulations with interference, we considered 3 different channel mod-els. First, we consider the case of having an ideal AWGN channel for bothusers, which also means that the interference experienced is the worst.
Figure 5.10 shows the outage after the first round of the HARQ protocol.The results show that the outage probability after the first round is around50%, which corresponds to the residual outage caused by the interference(the activity factor for the interference is set to 50%). In figure 5.11, weshow the spectral efficiency for the different MCS and PRB allocations. Wecan see the how, for different SNR values, there is a combination of MCS and
5.4 Performance Analysis of the Scheduler 85
−5 0 5 10 15 2010
−2
10−1
100
SNR [dB]
Pout1 (AWGN/AWGN)
mcs2
mcs3
mcs4
mcs5
mcs6
mcs7
mcs8
mcs9
mcs11
mcs12
mcs13
mcs16
mcs15
mcs18
activity factor=50%
Figure 5.10: Probability of Outage after the first round of the HARQprotocol. There is one interferer of the same strength as the user of interest
with a 50% probability of being active. The channel is AWGN for bothusers.
86
Chapte
r5
Pra
ctic
alSch
edule
rD
esig
nfo
rLT
EB
ase
Sta
tions
−5 0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
SNR [dB]
Spectral eff (AWGN/AWGN)
mcs2
mcs3
mcs4
mcs5
mcs6
mcs7
mcs8
mcs9
mcs11
mcs12
mcs13
mcs14
mcs15
mcs16
mcs18
activity factor =50%
Figure 5.11: Spectral efficiency for different resource allocations. There is one interferer of the same strength as the user ofinterest with a 50% probability of being active. The channel is AWGN for both users.
5.4 Performance Analysis of the Scheduler 87
−1 1 3 5 7 9 11 13 150
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
SNRdB
ratio of dimensions (ρ)
ρGaussian
ρsimulationAWGN/AWGN
interferenceactivity factor=50%
unconstrained modulation(PERFECT CE)
LTE coded−modulation(IMPERFECT CE)
Figure 5.12: Ratio of dimensions.
PRBs distribution that gives the maximum spectral efficiency. It should bementioned that some of the MCS values like MCS 17 are never used, sincethere is not an SNR region where it gives the highest spectral efficiency
We can compare this case to the one in the example with Gaussian signalsfrom section 4.4.2 in Chapter 4. For an outage constraint of 10−2, the ratioof dimensions is decreasing using every time, more dimensions in the secondround than in the first round. In the LTE case, the results show the sametrend, with a difference that can be explained by the fact that we do notassume perfect channel estimation and the constraints of the LTE coded-modulation against an unconstrained modulation represented by the mutualinformation expressions coming from the Gaussian expressions.
AWGN/Rayleigh
Next, we choose the channel of the user of interest as AWGN and a Rayleighfading channel for the interference. Figures 5.13 to 5.14 show the results forthis case. As we can see from figure 5.13, for the same latency constraintof 10−2, higher MCSs have higher probability of outage at the first round.And from figure 5.14, we observe that depending on the SNR, there is acombination of MCS and PRBs allocation that gives the highest spectralefficiency and, in general, more dimensions are used in the second round.This in contrast with the AWGN channels case.
88 Chapter 5 Practical Scheduler Design for LTE Base Stations
−5 0 5 10 15 20
100
SNR [dB]
Pout1 (AWGN/Rayleigh)
mcs2
mcs3
mcs4
mcs5
mcs6
mcs7
mcs8
mcs9
mcs11
mcs12
mcs13
mcs14
mcs15
mcs16
mcs18activity factor=50%
Figure 5.13: Probability of Outage after the first round of the HARQprotocol. There is one interferer of the same strength as the user of interestwith a 50% probability of being active. The channel is AWGN for the user
of interest and Rayleigh for the interferer.
5.4
Perfo
rmance
Analy
sisofth
eSch
edule
r89
−5 0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
SNR [dB]
Spectral eff (AWGN/Rayleigh)
mcs2
mcs3
mcs4
mcs5
mcs6
mcs7
mcs8
mcs9
mcs11
mcs12
mcs13
mcs14
mcs15
mcs16
mcs18
activity factor =50%
Figure 5.14: Spectral efficiency for different resource allocations. There is one interferer of the same strength as the user ofinterest with a 50% probability of being active. The channel is AWGN for the user of interest and Rayleigh for the interferer.
90 Chapter 5 Practical Scheduler Design for LTE Base Stations
−5 0 5 10 15 20
100
SNR [dB]
Pout1 (Rayleigh/Rayleigh)
mcs2
mcs3
mcs4
mcs5
mcs6
mcs7
mcs9
mcs12
mcs15
mcs18 activity factor=50%
Figure 5.15: Probability of Outage after the first round of the HARQprotocol. There is one interferer of the same strength as the user of interestwith a 50% probability of being active. Both the user of interest and the
interferer experience a Rayleigh fading channel.
Rayleigh/Rayleigh
Now, we look at the case of Rayleigh fading channel for both users. Inthis case, the outage probability after the first HARQ round (figure 5.15) ishigher because of the bad conditions of the channel. However, if we look atfigure 5.16, we see that not all of the MCS values are used for this case withlatency constraint of 1%. For example, MCS 4, 5, 12, have always a lowerspectral efficiency. This result may change for different latency constraints.
5.4
Perfo
rmance
Analy
sisofth
eSch
edule
r91
−5 0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
SNR [dB]
Spectral eff (Rayleigh/Rayleigh)
mcs2
mcs3
mcs4
mcs5
mcs6
mcs7
mcs9
mcs12
mcs15
mcs18
activity factor =50%
Figure 5.16: Spectral efficiency for different resource allocations. There is one interferer of the same strength as the user ofinterest with a 50% probability of being active. Both the user of interest and the interferer experience a Rayleigh fading
channel.
92 Chapter 5 Practical Scheduler Design for LTE Base Stations
EPA/AWGN
For a more realistic setup, we simulate a multipath EPA model, which isone of the LTE channel models [2], employed in an urban environment withsmall cell sizes (low delay spread). We model the channel of the interferer asAWGN since we assume that interference is coming from an indoors small-cell near the macro user. We show in figure 5.17 the outage probability afterthe first round of the HARQ protocol. In figure 5.18, we show the spectralefficiency. The performance is similar to the case with Rayleigh channel forthe user of interest with a high probability of outage and the best allocationcorresponding to the use of less resources in the first round for the high SNRregion. For the sake of visibility, we do not show all of the MCS values, butonly those that give the highest spectral efficiency, for different SNR regions.
−5 0 5 10 15 20
100
SNR [dB]
Pout1 (EPA/AWGN)
mcs3
mcs4
mcs6
mcs7
mcs9
mcs12
mcs15
mcs18activity factor=50%
Figure 5.17: Probability of Outage after the first round of the HARQprotocol. There is one interferer of the same strength as the user of interestwith a 50% probability of being active. The channel model is EPA for the
user of interest and AWGN for the interferer.
5.4
Perfo
rmance
Analy
sisofth
eSch
edule
r93
−5 0 5 10 15 200
0.2
0.4
0.6
0.8
1
1.2
1.4
SNR [dB]
Spectral eff (EPA/AWGN)
mcs3
mcs4
mcs6
mcs7
mcs9
mcs12
mcs15
mcs18
activity factor =50%
Figure 5.18: Spectral efficiency for different resource allocations. There is one interferer of the same strength as the user ofinterest with a 50% probability of being active. The channel model is EPA for the user of interest and Rayleigh for the
interferer.
94 Chapter 5 Practical Scheduler Design for LTE Base Stations
Figure 5.19: Openair Emulation Protocol Stack.
5.5 Scheduler under the Full LTE PHY/MAC Pro-
tocol Stack
The OAI platform provides a complete wireless protocol stack and radiohardware (see Figure 5.19) and comprises C-language implementations ofthe protocol stack and the PHY abstraction unit each corresponding to aparticular node in the network. At the network layer, OAI implements radioresource management, routing, multi-casting, topology control, and proxymobile IP.
To implement dynamic resource allocation policies such as we have de-scribed in Chapter 4 in real systems, changes have to be introduced in theMAC scheduling unit. These changes allows us to modify the MCS, andnumber of PRBs used across HARQ rounds. The allocation of resources hasto be signaled on the PDCCH with the use of the DCI every retransmission(retransmissions are always scheduled through PDCCH), instead of an auto-matic retransmission. The redundancy version has to be explicitly signaled,i.e. the redundancy version index (RVI) corresponding to the redundancyversion provided in the HARQ information. With respect to figure 5.19, thisinvolved updating the downlink scheduling policy in MAC Scheduling Unit(DLSCH). In addition, the LTE mechanisms for adapting the modulationorder was not available in the OAI implementation in neither the eNB orUE. This was added to both.
5.5 Scheduler under the Full LTE PHY/MAC Protocol Stack 95
5.5.1 Feasibility Evaluation
To test the feasibility of our resource allocation strategies on real LTEmodems, we evaluated the spectral efficiency of such shemes in a complete3GPP PHY/L2 implementation and compared it to the results obtained fromthe OAI unitary PHY link simulation. As a validation test, we consideredthe case without interference and AWGN channel in section 5.4.1.
Table 5.3, shows the comparison of the spectral efficiency performance ofthe PHY and PHY/MAC variable resource allocation implementations. Aswe can see, the results match those from the AWGN case in section 5.4.1. Tomake the comparison, we chose some SNR points and then run both casesfor 10000 frames.
The next step is to test our strategies under the presence of interference.For this purpose, in the next section we describe the scenario that we arecurrently investigating.
5.5.2 Interference Scenario Description
In this case, we consider the interference scenario of figure 5.20 where there isa single dominant interferer with sporadic traffic (eNB2) creating interferenceon the downlink of a macrocell user (UE1,1) who is also generating sporadictraffic. In this case, the interference can model a small cell (pico/femtocell)which is transmitting only a portion of the time.
To simulate the fully-loaded macrocell (eNB1), we add a second user(UE1,2) connected to the same eNB as the user of interest which is trans-mitting constant traffic. This user can be seen as a traffic generator used to“saturate” the traffic flow from eNB1. In a scenario with two small cells, thesecond UE (UE1,2) would have to be inactive.
In this scenario, the interference seen at the user of interest (UE1,1) isstrong while the interference coming from the macrocell on the small celluser is weak which can be due to the position of the small cell user (indoors)or the distance to the macrocell, etc. In our simulation, the interfering UEs
96 Chapter 5 Practical Scheduler Design for LTE Base Stations
Figure 5.20: Interference Scenario.
occupy artificially (by design) the same subframe which allows us to controlthe activity factor.
One target is the comparison of the distributed techniques with the ABSfeature (see section 2.2.1) introduced in LTE Release 10.
We choose the resource allocation based on statistics of the CQI thatthe eNB can keep over time and from this information we can infer, in thelong time, if the user is experiencing interference or not. Another way to getinformation about the state of the interference, would be to use an improvedfeedback assuming two possible SNR values to be fed back. One level ofSNR would correspond to the state with interference and the other to theno-interference state. Finally, we can use the expressions in section 4.4.2 tochoose the rate and dimensions to use in each of the transmission rounds.
To simulate an interference scenario, an activity factor can not be ex-plicitly simulated, however, we choose the load on the interference to be lowcompared to the user of interest. We are currently in the process of perform-ing a full protocol simulation with the allocation strategies in section 5.4.2with two UEs and two eNBs that corresponds to the downlink of the user ofinterest with one interferer. Table 5.4, shows the comparison of the spectral
Table 5.4: Interference results with PHY/MAC protocol stack
efficiency performance of the PHY and PHY/MAC variable resource alloca-tion implementations in the case of one interferer. As we can see, the resultsmatch those from the AWGN case in section 5.4.2. To make the comparison,
5.5 Scheduler under the Full LTE PHY/MAC Protocol Stack 97
we chose some SNR points and then run both cases for 10000 frames. We letthe interference to transmit on the same subframe as the user of interest. Ittransmits over all the PRBs to make sure there is interference. The resultsin table 5.4 are for the case when both users have the same strength.
98 Chapter 5 Practical Scheduler Design for LTE Base Stations
Chapter 6
Conclusions and Areas for
Further Research
6.1 Conclusions
We now summarize the main results of this thesis. In Chapter 3, we analyzedthe performance of small-cell networks, in particular, femtocell networks withinter-cell interference. We proposed a decentralized HARQ retransmissionprotocol that employs incremental redundancy combined with a receiver thatcan cancel two strong interferers. With the use of Monte Carlo simulationsto analyze the throughput, we showed that our scheme is effective at com-bating interference without requiring any coordination.
In Chapter 4, we first investigated the throughput of a point-to-point linkfor time-varying channels. By taking an information-theoretic approach, wederived the throughput of IR-HARQ with dynamic physical resource alloca-tion mechanisms. We presented rate-adaptation policies in the case of sparseand latency-constrained traffic. These policies can be applied for both down-link (DL) and uplink (UL) data. We then treated the case when outdatedCSI is available at the transmitter and we showed that even in the case ofoutdated information with a low correlation with the actual channel providesa gain in throughput.
We showed that, in general, by adapting the number of physical dimen-sions across rounds, we can exploit the interference mitigation effects ofHARQ using it not only to recover from errors but for interference can-cellation. We proposed efficient resource allocation algorithms to increase
99
100 Chapter 6 Conclusions and Areas for Further Research
the throughput, which can potentially come very close to optimal perfor-mance. Later in this chapter, we studied the case of interference networks,and we demonstrated the benefits of adapting the rate and physical dimen-sions across transmission rounds of HARQ protocols. We motivated the useof inter-round resource allocation with an example using Gaussian input sig-nals and we obtained a throughput higher than the ergodic capacity in thecase of zero-outage throughput and we showed that having an upper layerARQ in case of a residual outage probability results in a lower throughput.We studied practical scenarios (with signals coming from discrete constella-tions) without power control and CSI. In this case it is not possible to getzero-outage throughput, however, we benefit from the dynamic resource al-location and by imposing a constraint on the outage probability, we showedthat the throughput can be improved by varying the latency of the protocol.We also obtained results for the case of bursty interference in the UL channelwhen the interference is time-varying because is coming from other users. Inthis scenario, we use the activity factors to represent the probability of theinterference being active.
In the last part of this chapter, we showed some examples of practicalapplications for the analytical framework. We showed that adapting the re-sources across HARQ rounds brings a benefit for scenarios like a Manhattan-like topology or a macrocell overlaid by a femtocell. We ended the chapterwith a description of the procedure that has to be followed in performingPHY abstraction with the use of our information-theoretic quantities.
Finally, in Chapter 5, we presented the design of practical schedulersfor LTE base stations. We showed results from the implementation of ourresource allocation strategies in the OAI SDR platform. With the use of afully-compliant LTE modem implementation, we tackled the case of adaptingthe resource allocation under the constraints of LTE coded-modulation. Byusing the DCI, we adjusted the number of PRBs across HARQ transmissionrounds and we proved that by adapting the PRBs, we obtained a higherthroughput with respect to a static allocation. Moreover, we showed thatthe results are in agreement with those obtained from the theoretical analysisof Chapter 4. We also found that the performance of LTE codes differs fromthe theory since not all code rates behave in the same way. As a result ofthe rate-matching algorithm, higher MCS have a worse performance withrespect to lower MCS. The latter can be overcome, by lowering the MCS inthe consecutive rounds with the use of the LTE reserved MCS indexes forHARQ operation. In the last part, we presented a full PHY/MAC protocolstack implementation of the scheduling strategies.
6.2 Areas for further research 101
6.2 Areas for further research
We have concentrated on single antenna systems, however, our polices can bealso applied to MIMO systems. The framework can be extended to considersystems with multiple antennas both at the transmitter and receiver.
The model can also be extended to account for the scheduling of multipleUEs. In this case, the resources have to be allocated taking into account thenumber of active users and their QoS requirements. Priority-based policescan be also considered within this approach.
Our analytical framework can be used to derive the metrics needed toperform PHY abstraction. In order to perform large scale network perfor-mance evaluations, it is useful to abstract the PHY, since the simulationtime can grow exponentially and become computationally not feasible. PHYabstraction helps reducing the simulation time without a high computationalcost.
102 Chapter 6 Conclusions and Areas for Further Research
Appendices
103
Appendix A
Summary of the thesis in
French
A.1 Abstract en français
L’objectif de cette thèse est de concevoir, implémenter et évaluer les algo-rithmes cross-layer pratiques. Nous nous concentrons sur la technologie LTEet les réseaux non coordonnés post-LTE où l’interférence est un enjeu ma-jeur compte tenu des nouvelles tendances du trafic. L’objectif est d’allouerles ressources radio d’une manière efficace. Nous développons des modèlesd’interférence mathématiques et informatiques qui nous permettent com-prendre le comportement de ces réseaux et nous appliquons une approchebasée sur la théorie de l’information à différents scénarios d’interférence etcaractéristiques du trafic. Nous avons essayé de s’approcher le plus possiblede systèmes réels pour être en mesure de tester la faisabilité des techniquesproposées.
La thèse porte sur l’évaluation de la performance des scénarios avec in-terférence dans les réseaux 4G, en particulier celles qui découlent du dé-ploiements de cellules de petite taille (“small cell”). Le travail dans cettethèse s’adresse également à l’analyse de l’allocation des ressources et la re-quête de répétition automatique hybride (HARQ) à redondance incrémentalepour les interférences sporadiques (de façon plus générale les canaux vari-ables dans le temps) qui permet uniquement des informations partielles del’état du canal à l’émetteur. Ce travail est ensuite appliquée à la concep-tion d’ordonnanceur pour les stations de base LTE et inclut une analyse deperformance pour les modems LTE réels.
105
106 Appendix A Summary of the thesis in French
A.2 Introduction
Les systèmes de communication modernes, comme le Long Term Evolution(LTE) [6], exigent des taux élevés de données et une meilleure qualité deservice (QoS) de contrôle pour les services tels que la téléphonie vocale, lesjeux en ligne, la navigation Web, etc. A fin de faire face aux exigences de cesnouveaux types de services tout en offrant simultanément des débits élevés,les normes doivent évoluer et s’adapter. Tant les défis de la couche physique(PHY) et les exigences de qualité de service des applications doivent êtrepris en compte. Les futurs réseaux de communication sans fils requièrentune optimisation des paramètres dans toutes les couches. Dans une con-ception inter-couches, le taux, la puissance et le codage à la PHY peuventêtre adaptés pour répondre aux exigences des applications, compte tenu desconditions de canal(voir la figure A.1)
Figure A.1: Techniques Cross-Layer.
En élaborant des politiques qui combinent l’atténuation des interférencessur la couche PHY avec des algorithmes d’ordonnancement et de l’adaptationdu taux de la couche Medium Access Control (MAC), et en outre la gestiondes ressources radio à la couche de contrôle de liaison radio (RLC), nouspouvons obtenir un débit plus élevé, plus la bande passante et donc desréseaux plus efficaces. Dans le cadre de l’évolution de quatrième génération(4G) des systèmes, la mise en place de cellules de petite taille qui recouvrentle réseau cellulaire existant a été envisagée pour remplir les taches blanchesde couverture ou servir les utilisateurs mobiles et de plein air où le réseaucellulaire n’est pas déployée. Cependant, comme ces petites cellules ne peu-vent pas être connectés au réseau opérateur de raccordement, la coordinationentre eux pour la gestion des ressources est difficilement réalisable (voir la
A.2 Introduction 107
figure A.2).
Figure A.2: Femtocells sont un exemple de déploiements avec cellules depetite taille [4].
L’objectif de cette thèse est de concevoir, implémenter et évaluer les al-gorithmes cross-layer pratiques. Nous nous concentrons sur la technologieLTE et les réseaux non coordonnés post-LTE où l’interférence est un en-jeu majeur compte tenu des nouvelles tendances du trafic. L’objectif estd’allouer les ressources radio d’une manière efficace. Nous développons desmodèles d’interférence mathématiques et informatiques qui nous permettentcomprendre le comportement de ces réseaux et nous appliquons une approchebasée sur la théorie de l’information à différents scénarios d’interférence etcaractéristiques du trafic. Nous avons essayé de s’approcher le plus possiblede systèmes réels pour être en mesure de tester la faisabilité des techniquesproposées. Finalement, nous effectuons une étude de simulation complet quitermine par la implémentation et l’évaluation des techniques proposées dansla plate-forme de OpenAirInterface (OAI) [3].
A.2.1 Contributions et cadre de cette thèse
Le principal obstacle trouvée dans les réseaux de communication sans fils estle caractère variable dans le temps du canal physique. Compte tenu de cela,l’objectif d’un concepteur de système est de rendre le PHY/MAC intelligentspour simplifier la conception globale du réseau et optimiser les performances.Dans nos stratégies, nous avons essayé de prendre en compte les scénarios etles paramètres qui les rendent applicables aux systèmes pratiques.
Afin de mieux comprendre la pertinence de notre étude, nous commençonsau chapitre 2 en décrivant l’évolution des réseaux de communication sans fils,ainsi que les nouveaux scénarios avec interférence résultant de cette évolu-tion. Nous expliquons alors les fondements de réseaux LTE et interférence etnous décrivons également les bases de la programmation et de l’adaptationde liaison LTE.
Le Chapitre 3 traite de l’évaluation des performances des scénarios avec
108 Appendix A Summary of the thesis in French
interférence dans les réseaux 4G, en particulier celles qui découlent de dé-ploiements de femtocells. Dans ce chapitre , nous analysons le débit du réseauà l’aide d’un temps discret chaîne de Markov non homogène et quantités dethéorie de l’information de base. Nous considérons un ou deux brouilleursdominantes et nous étudions un système d’atténuation d’interférence décen-tralisée qui combine Répétez hybride automatique et demande ( HARQ ) etredondance incrémental (IR) avec un décodeur d’annulation d’interférence.Pour fins de comparaison , nous étudions également un système ARQ oùl’information n’est pas accumulée sur tours de transmission . Notre évalua-tion de la performance basée sur la modélisation analytique et l’évaluationde Monte Carlo de débit montre que notre système est efficace à la luttecontre l’interférence sans nécessiter de coordination . Les résultats ont étépubliés dans
• Villa, Tania; Merz, Ruben; Knopp, Raymond, “Interference man-agement in femtocell networks with Hybrid-ARQ and inter-ference cancellation”, in the proceedings of IEEE Asilomar Con-ference on Signals, Systems, and Computers, November 2011, PacificGrove, CA, USA.
Le quatrième chapitre commence par décrire quelques applications émer-gentes où notre analyse de l’information théorie peut être appliquée. Cechapitre est essentiel pour la compréhension de l’importance des systèmesde répartition des ressources dynamiques LTE. Contrairement à un tra-vail précédent, nous considérons l’allocation des ressources physiques nesont pas fixées dans les transmissions HARQ. Ce dernier est une possibilitéréelle dans ordonnanceurs pour les stations de base LTE (eNodeB ou ENB)et, au meilleur de notre connaissance, existe aucune méthode connue pourl’adaptation des ressources physiques sur tours HARQ lorsqu’il est soumis àdes canaux variant dans le temps, soit à cause de la décoloration ou variantdans le temps des interférences ou une combinaison des deux. En faisant cela,nous exploitons les atténuation des effets d’interférence de HARQ utilisantnon seulement de récupérer des erreurs mais pour annulation d’interférenceet nous proposons des algorithmes d’allocation des ressources efficaces pouraugmenter le débit, qui pourrait venir très proche de la performance opti-male.
Nous présentons d’abord une analyse des réseaux sans interférence avecles chaînes de variables dans le temps. Plutôt que d’effectuer des simulations,nous adoptons une approche théorique de l’information pour obtenir des ex-pressions analytiques qui représentent le débit à long terme d’une liaisonpoint-à-point et d’envisager des cas pratiques où il ya une contrainte sur laprobabilité d’interruption représentant la latence du protocole. Nous consid-érons les signaux d’entrée de Gauss et nous considérons le cas où l’indicateur
A.2 Introduction 109
de qualité de canal (CQI) rétroaction est indisponible, ou pas à jour.
Le quatrième chapitre examine ensuite le cas des réseaux d’interférences.Nous motivons l’utilisation de l’allocation des ressources inter-retour à traversune analyse simple mais illustratif avec des signaux de Gauss et les inter-férences. Nous incluons l’utilisation des facteurs d’activité qui modèle desfigures d’interférence sporadiques caractéristique de futurs déploiements deréseaux hétérogènes, en particulier l’ingérence vu des stations de base à cel-lules de petite taille avec du trafic en rafales dans le récepteur d’un utilisateurmacrocellulaire.
Enfin, nous examinons les scénarios d’interférence pratiques pour illus-trer les applications de notre cadre d’analyse. Nous modélisons une topologieManhattan comme ce qui représente un bloc d’appartements avec femtocellscréer des interférences. Nous modélisons alors un scénario macro-femto oùune macro-cellule est recouverte par une femtocell et nous regardons la liaisondescendante (DL) canal de l’utilisateur macrocellulaire lorsque l’interférenceprovient de la femtocell. Avec l’utilisation d’un facteur d’activité, nous mod-élisons le fait que la femtocell n’est pas active en permanence. Enfin, nous ex-pliquons la procédure à suivre pour effectuer PHY abstraction de l’utilisationde notre cadre d’analyse, étant donné l’importance de modéliser avec préci-sion les performances de liaison afin d’accélérer les simulations. Les résultatsont été publiés dans
• Villa, Tania; Merz, Ruben; Knopp, Raymond, “Adaptive modu-lation and coding with Hybrid-ARQ for latency-constrainednetworks”, in the proceedings of IEEE European Wireless Conference(EW2012), April 2012, Poznan, Poland.
• Villa, Tania; Merz, Ruben; Knopp, Raymond, “Adaptive transmis-sion and mutiple-access for sparse-traffic sources”, in the pro-ceedings of IEEE European Signal Processing Conference (EUSIPCO),August 2012, Bucharest Romania.
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation in heterogeneous networks”, in the proceedings of IEEEGlobal Communications Conference (GLOBECOM), December 2013,Atlanta, USA.
et a été accepté pour publication dans
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation for time-varying channels in next generation cellu-lar networks, Part I: a mathematical framework”, submitted toIEEE Transactions on wireless communications
110 Appendix A Summary of the thesis in French
et sera présenté comme une partie de
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation for time-varying channels in next generation cellu-lar networks, Part II: applications in LTE”, under preparation.
Le chapitre 5 traite avec la conception de planificateur pratique pour lesstations de base LTE . LTE offre une grande flexibilité en termes d’allocationdes ressources et, en particulier, les algorithmes d’allocation de ressourcespeut être adapté pour une classe particulière de trafic à des exigences spé-cifiques. Néanmoins, les travaux sont encore nécessaires pour exploiter effi-cacement cette flexibilité pour les applications émergentes. Dans ce chapitre,nous étudions les performances de nos politiques d’allocation de ressourcesdynamique pour la programmation des transmissions IR-HARQ sous les con-traintes de la LTE code-modulation. Dans ce chapitre, nous montrons lesrésultats de la implémentation de l’ordonnanceur dans le logiciel défini OAIradio ( SDR ) plate-forme [3] afin de tester les performances et la conformitéde nos stratégies d’allocation de ressources dans le LTE. Nous montrons quela implémentation de ces politiques dans un système réel est possible sansune coordination ou complexe et fastidieux processus d’optimisation. Nousmontrons que nos techniques d’ordonnancement travaillent pour différentsenvironnements et surtout, nous montrent que les résultats sont en accordavec les résultats théoriques présentés dans le chapitre 4. Les résultats serontprésentés comme une partie de
• Villa, Tania; Knopp, Raymond; Merz, Ruben, “Dynamic resourceallocation for time-varying channels in next generation cellu-lar networks, Part II: applications in LTE”, under preparation.
A.3 Résumé du Chapitre 2
A.3.1 Evolution des systemes san fils
L’évolution des systèmes de communication sans fil est principalement tiréepar l’introduction de nouveaux services et la disponibilité de technologiesplus avancées. Au cours des deux dernières décennies, les réseaux cellulairesont augmenté de façon exponentielle et la demande de services nouveaux etaméliorés est devenue un enjeu important pour les opérateurs. Il existe unbesoin pour de nouvelles technologies permettant de pallier les limitationsde capacité du réseau et de maintenir la qualité de service demandés par lesutilisateurs. Cela a motivé le développement de nouvelles normes comme(3GPP) la norme LTE [6] le projet de partenariat de troisième générationafin de fournir des débits plus élevés et une qualité de service améliorées dans
A.3 Résumé du Chapitre 2 111
Figure A.3: Macro-eNB vs pico-eNB.
les réseaux sans fil.
Le trafic de données va continuer à croître, avec des abonnements dedonnées mobiles et une augmentation du volume de données moyenne parabonnement. En fait, on s’attend à l’ensemble du trafic de données mobilesde poursuivre la tendance de doubler chaque année [1]. Cette croissance dutrafic et des services apportera de nouveaux défis techniques pour les opéra-teurs, les interférences étant l’une des meilleures performances de limitation.
A.3.2 Interférence dans les reseaux 4G
Pendant les 20 dernières années, il ya eu une augmentation massive du vol-ume de trafic, le nombre de périphériques connectés, et une demande accruepour les données vidéo. Les réseaux cellulaires futures devraient être enmesure de faire face à cette demande accrue et gérer tout le trafic d’unemanière efficace.
Il y a des nouveaux défis techniques et les scénarios possibles avec in-terférence qui varient selon le type de déploiements, les exigences, débit dedonnées élevé et les niveaux de qualité de service. Ces nouveaux scénariosd’interférence ont été considérés lors de la LTE sortie 10 normalisation:
• Interférence macro-picocell(voir la figure A.3).
• Interférence macro-Home-eNodeB (HeNB) (voir la figure A.4).
A.3.3 Gestion et adaptation de liaison pour LTE
Dans LTE, les algorithmes d’ordonnancement et de l’adaptation du taux dela couche MAC peuvent être combinés avec la gestion des ressources radio àla couche RLC pour obtenir un débit plus élevé, plus la bande passante etdonc des réseaux plus efficaces. La implémentation des algorithmes efficaces
112 Appendix A Summary of the thesis in French
Figure A.4: Macro-eNB vs HeNB.
pour la gestion des ressources radio, ordonnancement de paquets, le contrôled’admission ou de la puissance et de contrôle d’interférence sont importantspour optimiser la capacité et la performance.
La planification consiste à allouer les ressources de transmission, blocs deressources physiques (PRB) en LTE, pour les utilisateurs, toutes les possibil-ités de transmission. Compte tenu des variations rencontrées dans la qualitéd’un canal sans fils, le choix d’autres paramètres tels que la modulation etde codage (MCS) peut être adapté dans le but de maximiser la capacité dela cellule, tout en satisfaisant aux exigences en matière de qualité de servicede chaque utilisateur. De cette manière, le caractère aléatoire de la liaisonradio peut être prise en compte et exploité pour utiliser les ressources dela manière la plus efficace. Le planificateur interagit étroitement avec legestionnaire HARQ qui est responsable de la programmation des retrans-missions en cas de mauvaise réception. La norme LTE prend en charge laprogrammation dynamique, canal dépendant de renforcer la capacité globaledu système.
Dans LTE, le planificateur réside au eNB. La capacité est partagée entreplusieurs utilisateurs sur une base à la demande. Le but de l’ordonnanceurest de décider quel terminal ou une station de base d’émission et sur lequelensemble de ressources.
Semblable à des ordonnanceurs OFDMA utilisés sur le DL, les planifica-teurs SC-FDMA pour l’UL peuvent être à la fois du temps et de fréquenceopportuniste. Une différence importante entre DL et UL est que les rapportsCQI n’est pas nécessaire car le planificateur est situé à l’eNB qui permet demesurer la qualité du canal UL par signaux sonores de référence (SRS) [61].
En LTE, la bande passante disponible est divisée en N sous-porteuses.De N sous-porteuses, 12 ou 24 sous-porteuses adjacentes sont regroupées for-mant ce qu’on appelle un bloc de ressources (RB), qui représente la ressource
A.3 Résumé du Chapitre 2 113
Figure A.5: LTE resource grid [9].
de planification minimale pour UL et transmissions de DL et correspond à180 KHz de spectre (voir figure A.5). Les cadres LTE sont divisés en deuxfentes de durée Tslot = 0.5ms. Une fente est formée par NRB RB dans ledomaine de fréquence pour la durée de 6 ou 7 symboles OFDMA dans le do-maine du temps, en fonction de la longueur du préfixe cyclique (CP) utilisé.Le CP est utilisé à des fins de synchronisation et est fixée à chaque fente.Un sous-porteuse spécifique à l’intérieur de la RB est appelé un élément deressources (RE). Depuis les sous-porteuses en OFDMA sont orthogonalesn’y a aucune interférence de l’intérieur de la cellule, mais des interférencessurviennent dans les cellules voisines.
Le nombre de RBs disponibles dépend de la largeur de bande du canal(voir figure A.6), et en fonction de la longueur de la CP, un nombre dif-férent de symboles OFDMA est logé dans une rainure. La table A.1 donnele nombre différent de RBs disponibles pour chacune des largeurs de bandespécifiées dans la norme 3GPP avec le nombre correspondant de RB.
La gestion des ressources radio vise à la programmation des ressourcesdisponibles de la meilleure façon de permettre aux utilisateurs d’obtenirune qualité de service spécifiques. Un mécanisme intelligent doit considérerl’interférence créée avec des ressources physiques déjà attribués.
114 Appendix A Summary of the thesis in French
Figure A.6: LTE bandwidth et RBs [28].
Table A.1: NRB vs Downlink System Bandwidth
Bandwidth 1.4 MHz 3 MHz 5 MHz 10 MHz 15 MHz 20MHz
NRB 6 15 25 50 75 100
A.4 Résumé du Chapitre 3
L’expansion actuelle et croissante des réseaux cellulaires représente un défique l’aménagement et le déploiement nouvelle infrastructure est extrême-ment coûteux. Les cellules de petite taille sont considérées comme unesolution rentable pour étendre la couverture et réseaux de capacité de latechnologie LTE et post-LTE [4].
Les cellules de petite taille sont des points d’accès sans fils à faible puis-sance fonctionnant dans un spectre sous licence, utilisé en plein air pouraméliorer la couverture, ou à l’intérieur de l’entreprise ou de l’utilisation àdomicile. Le concept de cellules de petite taille comprennent les femtocells,pico et micro-cellules.
Dans le cas d’utilisation à domicile, les femtocells offrent une haute qual-ité, l’accès cellulaire à haut débit. Ils sont déployés par les utilisateurs finauxet connecté au réseau de l’opérateur en un numérique ligne d’abonné (DSL),un modem câble ou fibre optique [24]. En raison de la nature imprévue de dé-ploiements femtocell, ils peuvent souffrir de fortes interférences inter-celluleavec femtocells voisins dans les déploiements denses [26,54,68]. En outre, lacoordination est difficilement réalisable en raison des retards induits par leinfrastructure de backhaul de ces réseaux à domicile femtocell.
Dans ce chapitre, nous étudions un régime décentralisée d’atténuationd’interférence que combine IR-HARQ avec une annulation d’interférence dé-codeur. Notre évaluation de la performance basée sur la modélisation ana-
A.4 Résumé du Chapitre 3 115
lytique et expériences de Monte Carlo montre que notre système est efficacepour la lutte contre l’interférence sans nécessiter de coordination.
A.4.1 Interférence dans les reseaux small cells
Nous nous concentrons sur les technologies LTE et LTE-Advanced (dite4G) avec des couches physiques OFDMA et explorer des stratégies alter-natives pour limiter les interférences. OFDMA assure l’orthogonalité dessous-porteuses et, par conséquent, il n’y a pas de interférence intra-cellule.Cependant, l’interférence peut être vécue par les utilisateurs dans les cellulesadjacentes.
Contrairement au travail précédent, nous profitons de la nature non-gaussien d’ingérence dans les déploiements femtocell où il ya généralement unseul ou deux fortes brouilleurs dominantes [60]. Nous considérons des signauxprovenant d’alphabets discrètes afin de pouvoir bénéficier de la structure del’interférence. Les signaux gaussiens atteindre le maximum d’efficacité spec-trale. Cependant, les systèmes pratiques utilisent des petits entrée alphabetsfinis de taille.
Sous ces hypothèses, nous proposons une stratégie décentralisée qui com-bine l’annulation d’interférence décodage [32] avec une politique de redon-dance HARQ incrémentale [20].
Pour évaluer la performance de cette stratégie, nous développons unemodèle analytique du débit obtenu par un protocole HARQ avec une an-nulation d’interférence décodeur. En particulier, notre modèle s’appuie surune caractérisation de l’information-théorique du taux réalisable avec uneannulation d’interférence.
A.4.2 Modèle du system
Nous nous concentrons sur un scénario de liaison descendante. Sans perte degénéralité, nous considérons actuellement de l’antenne unique. Par ailleurs,les transmissions sont fendue et parfaitement synchronisés
Nous avons Nu émetteurs, où noeud 0 est l’émetteur de intérêt et lesautres Nu − 1 émetteurs sont brouilleurs. Nous laissons dk soit la distanceentre le noeud k et le récepteur. Les systèmes 4G sont basées sur une couchephysique OFDMA [61]. Nous laissons y[m] le signal reçu dans un RB parti-culier au moment de m. Dans LTE, chaque RB est défini comme un groupede sous-porteuses K. Au sein d’une cellule donnée, on suppose que les RBssont orthogonales entre elles. Par conséquent, nous pouvons écrire le signal
116 Appendix A Summary of the thesis in French
reçu y[m] au temps m comme
y[m]=K−1∑
j=0
Nu−1∑
k=0
√
Pkd−αk hk,j [m]µk,j [m]xk,j [m] + zj [m]. (A.1)
où xk,j [m] est le signal transmis à partir du noeud k dans le jième sous-porteuse d’un RB particulier, µk, j[m] est un soi-disant de facteur d’activité,zj [m] est le bruit thermique, Pk est la puissance de transmission, α estl’exposant de perte de trajet et hk,j [m] est le coefficient du canal. Nousmodélisons zj [m] comme le bruit, un processus indépendamment et iden-tiquement distribuées (iid) de moyenne nulle blanc gaussien additif (AWGN)de variance σ2.
Si nous nous concentrons sur une sous-porteuse particulier, le signal reçudans le jième sous-porteuse au temps m est
yj [m] =
Nu−1∑
k=0
√
Pkd−αk hk,j [m]µk,jxk,j [m] + zj [m] (A.2)
La variable aléatoire hk,j [m] est iid pour chaque emplacement avec unedistribution de Rayleigh. Par conséquent, le coefficient de canal reste con-stante pendant la durée d’un créneau. Les facteur d’activité modèlisent lacharge de trafic et/ou de transmission discontinue (DTX) des caractéristiquesde la technologie LTE [61]. Nous modélisons µk,j avec un Bernoulli iid, unedistribution avec le paramètre p.
Ces caractéristiques sont prises en compte car elles ont un effet direct surla distribution d’interférence. Le protocole de retransmission est un schémaHARQ utilisant IR [20]. Pour des fins de comparaison, nous considéronségalement une schéma ARQ simple qui retransmet le même bloc de don-nées dans le cas de transmission infructueuse. Le paramètre Mmax est lenombre maximum de tours ARQ. Par conséquent, une trame donnée peutêtre retransmis à la plupart des Mmax fois et est jeté si Mmax est atteint.Nous supposons parfaite information de l’état du canal(CSI) et des signauxd’interférence désirés au niveau du récepteur et nous permettent R de définirla vitesse de transmission donnée par un MCS particulier.
A.5 Résumé du Chapitre 4
Les performances de la technologie LTE en termes d’efficacité spectrale etdes débits de données disponibles est, relativement parlant, plus limitée par
A.5 Résumé du Chapitre 4 117
l’interférence de cellules adjacentes que par rapport aux normes de com-munication précédentes [27]. Moyens pour réduire ou contrôler l’interférenceinter-cellules pourrait avoir des retombées importantes pour les performancesLTE, notamment en termes de la qualité de service fourni à chaque utilisa-teur.
Les algorithmes d’allocation des ressources efficaces sont l’un des élé-ments clés dans la fourniture de haute efficacité du spectre LTE [34]. Unalgorithme d’ordonnancement intelligent peut, en même temps, contribuer àréduire l’impact des interférences.
La performance d’un algorithme d’ordonnancement particulier peut êtreévaluée à l’aide de simulations exhaustives, mais il est temps et a un coût decalcul élevé. En utilisant la théorie de l’information, nous pouvons analyserle débit possible, ou de l’efficacité spectrale, sous différentes hypothèses dumodèle de système.
Plutôt que des simulations, nous adoptons une approche théorique del’information pour obtenir des expressions analytiques qui représentent ledébit à long terme du réseau et d’envisager des cas pratiques où il ya unecontrainte sur la probabilité d’interruption représentant la latence du pro-tocole. Nous considérons les signaux provenant des alphabets discrets pourmodéliser des systèmes pratiques et nous considérons le cas où les informa-tions de CQI sont indisponibles, ou pas à jour.
A.5.1 Applications clés
Les réseaux sans fils cellulaires peuvent être divisés en homogènes et HetNets.Hétérogène implique qu’il existe différents types de cellules dans le réseau,à savoir les stations de base de faible puissance sont distribués à traversd’un réseau de cellules macro. Ces stations de base de faible puissance peu-vent être microcells, les picocells, relais, femtocells ou systèmes d’antennesdistribuées [33]. D’une part, les microcells, les picocells et les relais sontdéployées par l’opérateur pour augmenter la capacité et la couverture dansles lieux publics, les entreprises de bâtiments, etc. D’autre part, les fem-tocells sont déployées à la maison pour améliorer la capacité déployée parl’utilisateur. On note généralement les stations de base de faible puissancepar de cellules de petite taille.
Les HetNets cellulaires fonctionnent généralement sur spectre sous licencedétenue par l’opérateur de réseau. L’ingérence plus grave est connu lorsqueles cellules de petite taille sont déployés sur la même fréquence porteuseque les macro-cells [47]. Scénarios de brouillage plus difficiles sont identifiésdont l’interférence peut venir à travers les couches (macro–à cellules de pe-
118 Appendix A Summary of the thesis in French
(a) Interference scenarios DL.
(b) Interference scenarios UL.
Figure A.7: Figure (a) montre les scénarios d’interférence pour HetNetsdans le DL, la figure (b) montre les scénarios d’interférence pour HetNets
dans le UL
tite taille, cellules de petite taille–macrocells), par exemple, un utilisateurmacrocellulaire loin de la station de base transmet à un très haut pouvoir deblesser les femtocells dans les environs . Les interférences peuvent aussi êtrevécue entre les petites cellules à la fois dans l’UL et canaux de DL (voir lafigure A.7). Dans le cas d’interférences inter-couches, le planificateur de lamacro-cell doit prendre en compte l’interférence sporadique des femtocells,car ils seront servent seulement quelques utilisateurs.
La communication M2M, une partie de la révolution d’internet des objets(IoT), devrait permettre de créer un nombre croissant d’appareils connec-tés, qui dépassera les communications homme-à-homme au cours des annéessuivantes (50 milliards de machines contre sept milliards de personnes pour2011) [25,53].
Dans les scénarios de trafic épars et latence limitées, les arrivées de pa-
A.5 Résumé du Chapitre 4 119
Figure A.8: Le trafic épars dans un scénario de retard limité. Les arrivéesdu traffic dans la couche MAC eNB sont rares comme représenté en bleu (il
y a trois d’entre eux). La contrainte de latence est de quatreemplacements, à savoir il ăăăă sont jusqu’à quatre attributions de canaux
PDSCH possibles. En raison de la circulation clairsemée, le CQI n’est pas àjour ou n’est pas disponible sur la première tranche.
quets sont sporadiques et doivent être programmés sous une contrainte delatence (voir la figure A.8). Dans ce contexte, le CQI est généralementobsolète ou indisponible. Notez que le CQI pas à jour se produit aussi àcause de mobilité modérée à élevée, de l’insuffisance de liaison montanteCQI périodicité ou de non-stationnaire interférence inter-cellules. Celui-cideviendra de plus en plus importante avec LTE Release 10 réseaux et leurhétérogénéité inhérente. Par conséquent, le planificateur doit fonctionner àl’aveuglette pour AMC et ne peut bénéficier de feedback après le premiertour de transmission HARQ sous forme de signalisation ACK/NACK.
A.5.2 Analysis pour les réseaux avec interférence
Nous considérons un système de transmission à fentes et nous adoptons uneapproche de l’information-théorique pour analyser les performances de débit.Quand il y a plus d’un utilisateur, nous supposons que toutes les trans-missions à chaque emplacement sont synchronisées et nous faisons le pro-cessus d’interférence aléatoire avec l’utilisation de facteurs d’activité. Lesderniers modèlisent des figures d’interférence sporadiques caractéristiquesdes futurs déploiements des réseaux hétérogènes, en particulier l’ingérencevu des stations de base à petites cellules avec du trafic en rafales dans lerécepteur d’un utilisateur macrocellulaire. Ils peuvent aussi modéliser lesréseaux bi-porteuses avec la programmation “cross-carrier”. Dans ce type deréseau, nous pouvons parler de transporteurs propres et sales. D’une part,les transporteurs propres sont utilisés par la macro-cell pour transporterleurs données ainsi que la signalisation pour les petites cellules en raison deleur propriété d’interférence contrôlée. D’autre part, les transporteurs sales
120 Appendix A Summary of the thesis in French
Figure A.9: Modèle de codage
interfèrent transporteurs où la “nettoyage” se fait avec l’utilisation de HARQ.
Nous considérons un maximum de Mmax tours de transmission HARQ etle canal est iid ou constante sur toutes les manches du protocole de transmis-sion. Après chaque transmission, nous recevons un accusé de réception sanserreur (ACK ou NACK) indiquant une transmission réussie ou non. Nousdéfinissons la probabilité de panne comme étant infructueuse de recevoir cor-rectement les informations à la fin du protocole HARQ. Cette probabilité setraduit par la latence du protocole et la qualité de service dans notre système.
En général, nous définissons Rr comme le taux de code à l’reme round.Pour un utilisateur particulier, nous définissons le nombre de dimensionsdans le temps comme Tdim et le nombre de dimensions de la fréquence commeLr. Soit L′r le nombre de dimensions de la fréquence à la tour r. Ensuite, àchaque tour de transmission, le nombre total de dimensions est L′rTdim. Ensupposant que le canal ne varie pas au cours de Tdim dimensions de tempset pour une longueur de paquet d’information de B bits, le taux Rr au remeronde, en bits/dim est donnée par :
Rr =log2B
L′rTdim
bits/dim. (A.3)
Dans IR-HARQ, la retransmission comprend le même ensemble de bitsd’information que l’original, cependant, l’ensemble de bits codés sont choi-sis différemment, et ils peuvent contenir des bits de parité supplémentaires.Dans chacun de la transmission arrondit il y a LrTdim dimensions, cepen-dant, ce nombre n’est pas nécessairement le même partout tours selon lanorme LTE [7] (voir la figure A.9).
Dans le contexte de LTE, le nombre de dimensions physiques LrTdim
désigne le nombre de blocs de ressources attribués à un utilisateur dansune sous-trame de durée 1ms soit un temps de transmission d’intervalle(TTI). Il existe dans la plupart des deux blocs de transport fournis à la
A.6 Résumé du Chapitre 5 121
couche physique dans le cas de multiplexage spatial [28]. Dans un systèmeLTE mono-utilisateur, il y a seulement un bloc de transport dans un TTI,représentant un seul mot de code “dans l’air” en même temps. Chaque blocde transport est effectuée par un processus HARQ, et chaque processus estassocié à une sous-trame (nombre de processus est fixe). Dans notre modèle,si le nombre de dimensions d’un utilisateur est inférieur au nombre maxi-mum de ressources disponibles NT ,(LrTdim < NT ), alors le reste ne sera pasutilisée. Bien que n’étant pas possible dans la norme LTE actuelle, nouspourrions proposer d’utiliser les ressources non utilisées pour transmettre demultiples mots de code en parallèle (en même temps), pour augmenter ledébit. Dans un système multi-utilisateur, les dimensions restantes seraientaffectées à d’autres utilisateurs, et donc l’efficacité du protocole doivent êtrechoisis afin de maximiser l’efficacité spectrale globale de la cellule.
Pour modéliser des systèmes pratiques, nous dérivons des expressionspour l’information mutuelle en supposant constellations discrètes. Nousciblons des réseaux LTE Release 10 d’une couche physique OFDMA, et nousétudions la fois la mono-utilisateur et un cas de brouilleurs dominantes.
Dans ce chapitre, nous avons démontré les avantages de l’adaptationdu taux et les dimensions physiques travers tours de protocoles HARQ detransmission. Nous avons obtenu un débit plus élevé que la capacité er-godique dans le cas d’indisponibilité de débit zéro et nous avons montré quela présence d’un message ARQ de la couche supérieure dans le cas d’ uneprobabilité résiduel de panne il y a un débit inférieur. Dans les scénariospratiques sans contrôle de puissance et de l’information d’état de canal, iln’est pas possible d’obtenir zéro panne débit. Toutefois, nous bénéficions del’allocation dynamique des ressources et en imposant une contrainte sur laprobabilité d’interruption, nous pouvons améliorer le débit en faisant varier letemps de latence du protocole. En motivant l’utilisation de l’allocation desressources inter-ronde, nous avons dérivé politiques distribués d’allocationdes ressources qui sont applicables pour l’UL et canaux de DL. Nous avonsconsidéré les interférences qui peuvent être sporadique en raison des carac-téristiques de déploiements de HetNets. Plutôt que d’effectuer des simula-tions, nous avons dérivé les expressions analytiques, basée sur la modélisationde l’information mutuelle, qui captent la performance de débit du réseau,avec ou sans contrainte de latence.
A.6 Résumé du Chapitre 5
Un élément clé de la technologie LTE est l’adoption de procédures de gestiondes ressources radio de pointe afin d’accroître les performances du système.L’utilisation efficace des ressources radio est essentielle pour atteindre les
122 Appendix A Summary of the thesis in French
objectifs de performance du système [21]. Dans ce chapitre, nous étudionsles performances de nos politiques d’allocation de ressources pour la planifi-cation des transmissions HARQ sous les contraintes de la modulation codéde LTE.
L’algorithme HARQ et l’adaptation de débit est mis en place et géré parle programmateur à la couche MAC. Le mécanisme de planification joue unrôle fondamental, car il est responsable du choix, avec bien du temps et dela fréquence granularité, comment distribuer les ressources radio entre lesdifférents utilisateurs, en tenant compte des conditions de canal actuels etexigences de QoS. L’algorithme d’ordonnancement est une question de la sta-tion de base, il n’est pas spécifié dans la norme. La norme ne donne pas desmesures/rapports de la qualité du canal et des procédures spécifiques pourl’allocation dynamique des ressources [28] nécessaire pour la implémentationdu terminal (et par conséquent de la station de base ainsi). Ceci permetd’avoir des algorithmes spécifiques au fournisseur qui peuvent être optimiséspour des scénarios spécifiques.
La conception des mécanismes d’allocation des ressources efficaces de-vient cruciale pour un fonctionnement efficace du réseau. Une complex-ité faible et l’évolutivité sont des exigences fondamentales pour trouver lameilleure décision d’attribution. Problèmes d’optimisation complexes et nonlinéaires ou de recherche exhaustive sur toutes les combinaisons possiblesserait trop coûteux en termes de coût et de temps de calcul [45]. Un al-gorithme doit être facilement mis en place et, surtout, devrait exiger trèsfaible coût de calcul. Plusieurs solutions théoriques peuvent être trouvéesdans la littérature, mais quand on les étudie de près, ils ne peuvent êtredéployés dans des systèmes réels compte tenu de la difficulté à mis en placedans des dispositifs réels et le coût de calcul élevé requis. Stratégies ro-bustes devraient garantir la possibilité de travailler dans des scénarios trèsdifférents. Il ne devrait pas nécessiter des réglages de paramètres forts, ou ildevrait au moins dynamiquement adapter ces paramètres aux changementsenvironnementaux [21].
A.6.1 Implémentation sur OpenAirInterface
OAI est une plate-forme de développement matériel/logiciel open-source etopen-forum pour l’innovation dans le domaine des communications radionumériques [3]. Il a été créé par le ministère des communications mobilesà EURECOM sur la base de son expérience en financement public de R&Deffectuée dans le cadre de projets de recherche en collaboration [15]. Un desavantages les plus importants d’une plate-forme complète de SDR est que lemême code peut être utilisé pour l’émulation comme dans une implémenta-tion réelle, d’assurer une transition en douceur entre les simulations pour les
A.6 Résumé du Chapitre 5 123
Figure A.10: Pile de protocole pour l’emulation Openair.
tests réels.
La Plate-forme OAI fournit une pile de protocole sans fil complète etmatériel de radio (voir la figure A.10) et comprend les implémentations enlangage C de la pile de protocole et l’unité PHY d’abstraction correspondantchacune à un node particulier dans le réseau. À la couche réseau, l’OAI misen place la gestion des ressources radio, le routage multi-casting, contrôle detopologie, et proxy IP mobile.
Pour mettre en place des politiques d’allocation dynamique des ressourcestelles que nous l’avons décrit dans le chapitre 4 dans les systèmes réels, deschangements doivent être introduits dans l’unité de planification MAC. Ceschangements nous permettent de modifier le MCS, et le nombre de RBs util-isés dans les tours HARQ. L’allocation des ressources doit être signalé surle PDCCH avec l’utilisation de la DCI chaque retransmission (les retrans-missions sont toujours prévues par PDCCH), au lieu d’une retransmissionautomatique. La version de redondance doit être explicitement signalé, àsavoir l’indice de version de redondance (RVI) correspondant à la versionde redondance dans les informations fournies HARQ. En ce qui concerne lafigure A.10, il s’agissait de mettre à jour la politique d’ordonnancement deliaison descendante dans l’unité de planification MAC (DLSCH). En outre,les mécanismes de LTE pour l’adaptation de l’ordre de modulation n’étaientpas disponibles dans la implémentation de l’OAI ni dans la eNB ou UE. Cecia été ajouté à la fois.
124 Appendix A Summary of the thesis in French
A.6.2 Application des techniques pour les modems LTE
Un algorithme d’ordonnancement répond à une métrique pré-formulé re-latif à la capacité qui est optimisée dans toutes les solutions d’allocationde ressources possibles satisfaisant un ensemble d’exigences prédéterminéestelles que la qualité de service, efficacité spectrale, ou de latence. Pour cal-culer l’efficacité spectrale, nous devons tenir compte de la TBS, l’ordre demodulation et la probabilité de panne:
Speceff =Qm,1
G1TBS
(
1−Pout,1
)
+
(
Qm,1
G1+
Qm,2
G2
)
TBS(
Pout,1(1−Pout,2))
(A.4)où Gi est le nombre de bits codés par mot de code, et Qm,i l’ordre de mod-ulation utilisé à l’ieme ronde. Pout,1 est la probabilité d’interruption aprèsle premier tour, et Pout,2 est la probabilité d’interruption après le deuxièmetour.
A titre d’exemple de l’optimisation de (A.4), nous fixons le TBS pourtoutes les allocations possibles à environ 1000 bits et de la table 7.1.7.2.1-1dans [7], nous obtenons l’indice du TBS en fonction du nombre de PRBsutilisés dans le premier tour et le TBS plus proche de 1000. Enfin, le tableau7.1.7.1-1 (voir [7]) nous donne l’indice et de modulation afin MCS correspon-dante.
Table A.2: Allocatin des ressources vs Modulation et tailles de bloc detransport
La table A.2 montre les différentes allocations utilisées pour générer lesrésultats en chiffres 5.2 et 5.3, où
(
NDLRB
)
1représente le nombre de PRBs al-
louées dans le premier tour. Nous utilisons ce tableau pour tester différentesallocations et étudier la performance des codes LTE. En utilisant la DCI,
A.6 Résumé du Chapitre 5 125
nous pouvons signaler les nouvelles informations par rapport au nombre dePRBs dans les tours du protocole HARQ consécutives. Le MCS peut égale-ment être adapté, mais le TBS reste fixe.
Pour tester la faisabilité de nos stratégies d’allocation de ressources surde véritables modems LTE, nous avons évalué l’efficacité spectrale de cessytèmes dans une implémentation complète 3GPP PHY/L2 et comparé lesrésultats obtenus à partir de la simulation unitaire de liaison PHY OAI.Comme un test de validation, nous avons considéré le cas sans interférenceet le canal AWGN dans la section 5.4.1.
Pour le scénario avec interférence de la figure A.11 où il y a un seulbrouilleur dominant avec un trafic sporadique (eNB2) qui crée des inter-férences sur la liaison descendante d’un utilisateur macrocellulaire (UE1,1)qui est également la génération de trafic sporadique. Dans ce cas, l’interférencepeut modéliser une petite cellule (pico/femtocell) qui transmet seulement unepartie du temps.
Pour simuler la macro-cellule entièrement chargée (eNB1), nous ajoutonsun deuxième utilisateur (UE1,2) connecté à la même eNB que l’utilisateurd’intérêt qui transmet le trafic constant. Cet utilisateur peut être considérécomme un générateur de trafic utilisé pour “saturer” le flux de eNB1 de cir-culation. Dans un scénario avec deux petites cellules, le second UE (UE1,2)devrait être inactif.
Dans ce scénario, l’interférence vu à l’utilisateur d’intérêt (UE1,1) estsolide pendant l’interférence provenant de la macro-cellule sur l’utilisateurde la petite cellule est faible, ce qui peut être dû à la position de l ’util-isateur de la petite cellule (à l’intérieur) ou la distance à la macro-cellule,etc. Dans notre simulation, les UE interférents occupent artificiellement (parconception) de la même sous-trame qui nous permet de contrôler le facteurd’activité.
Une cible est la comparaison des techniques distribués avec la fonctionABS (voir la section 2.2.1) mis en place en LTE sortie 10.
Nous choisissons l’allocation des ressources basée sur les statistiques del’CQI que le eNB peut garder au fil du temps et à partir de cette infor-mation, nous pouvons déduire, dans le temps, si l’utilisateur connaît lesinterférences ou non. Une autre façon d’obtenir des informations sur l’étatde l’interférence, serait d’utiliser un retour parmi les meilleurs en supposantdeux valeurs possibles de SNR à être réinjectées. Un niveau de SNR cor-respondrait à l’état de l’interférence et l’autre à l’état de non-interférence.Enfin, nous pouvons utiliser les expressions dans la section 4.4.2 pour choisirle taux et les dimensions à utiliser dans chacun des tours de transmission.
126 Appendix A Summary of the thesis in French
Figure A.11: Scénario avec interférence.
Pour simuler un scénario d’interférence, un facteur d’activité ne peut pasêtre simulée de manière explicite, toutefois, nous choisissons la charge surl’interférence d’être faible par rapport à l’utilisateur d’intérêt. Nous sommesactuellement dans le processus de réalisation d’une simulation de protocolecomplet avec les stratégies de répartition dans la section 5.4.2 avec deux UEet deux eNB qui correspond à la liaison descendante de l’utilisateur d’intérêtavec un brouilleur.
A.7 Conclusion
Nous résumons maintenant les principaux résultats de cette thèse. Dans lechapitre 3, nous avons analysé la performance des réseaux à petites cellules,en particulier les réseaux femtocell avec interférence inter-cellules. Nousavons proposé un protocole de retransmission HARQ décentralisée qui em-ploie redondance incrémentale combiné avec un récepteur qui peut annulerdeux de fortes interférences. Avec l’utilisation de simulations de Monte Carlopour analyser le débit, nous avons montré que notre système est efficace àla lutte contre l’interférence sans nécessiter de coordination.
Dans le chapitre 4, nous avons d’abord étudié le débit d’une liaison point-à-point pour les canaux variant dans le temps. En adoptant une approchede l’information théorique, nous avons calculé le débit de l’IR-HARQ avecles mécanismes d’allocation dynamique de ressources physiques. Nous avonsprésenté les politiques taux d’adaptation dans le cas du trafic sporadiqueet la latence limitées. Ces stratégies peuvent être appliquées à la fois pourla liaison descendante (DL) et la liaison montante (UL) des données. Nousavons ensuite traité le cas où CSI à jour est disponible à l’émetteur et nousavons montré que, même dans le cas de l’information pas à jour avec unefaible corrélation avec le canal actuel permet un gain de débit.
A.7 Conclusion 127
Nous avons montré que, en général, en adaptant le nombre de dimensionsphysiques à travers des tours, nous pouvons exploiter les effets d’atténuationd’interférence de HARQ en l’utilisant non seulement pour récupérer des er-reurs mais pour annulation d’interférence. Nous avons proposé des algo-rithmes d’allocation des ressources efficaces pour augmenter le débit, quipourrait venir très proche de la performance optimale. Plus tard dans cechapitre, nous avons étudié le cas des réseaux d’interférence, et nous avonsdémontré les avantages de l’adaptation du taux et les dimensions physiquesa travers des tours de protocoles HARQ de transmission.
Nous motivés à l’utilisation de l’allocation des ressources inter-ronde avecun exemple en utilisant des signaux d’entrée de Gauss et nous avons obtenuun débit plus élevé que la capacité ergodique dans le cas de zéro panne débitet nous avons montré que la présence d’une couche ARQ supérieur en casde panne résiduelle résultats de probabilité à un débit inférieur. Nous avonsétudié des scénarios pratiques (avec des signaux provenant de constellationsdiscrètes) sans contrôle de puissance et CSI. Dans ce cas, il n’est pas pos-sible d’obtenir zéro panne débit, cependant, nous bénéficions de l’allocationdynamique des ressources et en imposant une contrainte sur la probabil-ité d’interruption, nous avons montré que le débit peut être améliorée enfaisant varier le temps de latence du protocole. Nous avons également obtenudes résultats pour le cas de brouillage sporadique dans le canal UL lorsquel’interférence est variable dans le temps en raison vient des autres utilisa-teurs. Dans ce scénario, on utilise les facteurs d’activité pour représenter laprobabilité de l’interférence étant actif.
Dans la dernière partie de ce chapitre, nous avons montré quelques exem-ples d’applications pratiques pour le cadre analytique. Nous avons montréque l’adaptation des ressources entre les tours HARQ apporte un avantagepour les scénarios comme une topologie Manhattan ou comme une macro-cellule ou recouvert par une femtocell. Nous avons terminé le chapitre parune description de la procédure qui doit être suivie dans l’exécution du PHYabstraction de l’utilisation de nos quantités d’information-théorique.
Dans le chapitre 5, nous avons présenté la conception d’ordonnanceurspratiques pour les stations de base LTE. Nous avons montré les résultats dela implémentation de nos stratégies d’allocation de ressources dans la plate-forme SDR OAI.
Avec l’utilisation d’une application de modem LTE entièrement conforme,nous avons abordé le cas de l’adaptation de l’allocation des ressources dansle cadre des contraintes de LTE code-modulation. En utilisant la DCI, nousavons ajusté le nombre de PRBs à travers des tours de transmission HARQ
128 Appendix A Summary of the thesis in French
et nous avons prouvé que par l’adaptation des PRBs, nous avons obtenuun débit plus élevé par rapport à une allocation statique. De plus, nousavons montré que les résultats sont en accord avec ceux obtenus à partirde l’analyse théorique du chapitre 4. Nous avons également constaté queles performances des codes LTE est différente de la théorie, comme tous lesdébits de code se comportent de la même façon. À la suite de l’algorithmetaux d’appariement, MCS supérieur ont une moins bonne performance parrapport à MCS inférieure. Ce dernier peut être surmonté, par l’abaissementde la MCS dans les tours consécutifs à l’utilisation de la technologie LTEréservé indices de MCS pour le fonctionnement HARQ. Dans la dernière par-tie, nous avons présenté un implémentation PHY/MAC complète de la pilede protocole des stratégies d’ordonnancement.
Notre cadre d’analyse peut être utilisée pour calculer les paramètresnécessaires pour effectuer PHY abstraction. Afin de procéder à des éval-uations de performance des réseaux à grande échelle, il est utile de faireabstraction de la PHY, depuis le temps de simulation peut croître de façonexponentielle et devenir des calculs pas possible. La PHY abstraction aide àréduire le temps de simulation sans coût de calcul élevé.
Bibliography
[1] Ericcson mobility report, june 2013.
[2] http://www.3gpp.org/ftp/specs/archive/ (last visited: 3 october 2009).
[3] http://www.openairinterface.org/ (last visited: 25 july 2013).
[4] http://www.smallcellforum.org/ (last visited: 23 july 2013).
[5] Ericsson white paper: Lte release 12, January 2013.
[6] 3GPP TR 36.201 3rd Generation Partnership Project. Technical spec-ification group radio access network; evolved universal terrestrial radioaccess (e-utra); lte physical layer; general description, March 2011.
[7] 3GPP TR 36.213 3rd Generation Partnership Project. Technical spec-ification group radio access network; evolved terrestrial radio access(e-utra); physical layer procedures, March 2011.
[8] 3GPP TS 36.101 3rd Generation Partnership Project. Technical spec-ification group radio access network; evolved terrestrial radio access(e-utra); user equipment (ue) radio transmission and reception, March2009.
[9] 3rd Generation Partnership Project, 3GPP TS 36.211. Technical spec-ification group radio access network; evolved universal terrestrial radioaccess (E-UTRA); physical channels and modulation, December 2008.
[10] A. Annamalai and C. Tellambura. A simple exponential integral repre-sentation of the generalized marcum q-function qm (a, b) for real-orderm with applications. In Military Communications Conference, 2008.MILCOM 2008. IEEE, pages 1 –7, nov. 2008.
[11] Huseyin Arslan. Cognitive Radio, Software Defined Radio, and AdaptiveWireless Systems (Signals and Communication Technology). Springer-Verlag New York, Inc., Secaucus, NJ, USA, 2007.
[12] F. Baccelli and P. Brémaud. Palm Probabilities and Stationary Queues.Springer Verlag Lecture Notes in Statistics, March 1987.
129
130 Bibliography
[13] K. Balachandran, S.R. Kadaba, and S. Nanda. Channel quality esti-mation and rate adaptation for cellular mobile radio. Selected Areas inCommunications, IEEE Journal on, 17(7):1244 –1256, jul 1999.
[14] Hari Balakrishnan, Peter Iannucci, Jonathan Perry, and DevavratShah. De-randomizing shannon: The design and analysis of a capacity-achieving rateless code. CoRR, abs/1206.0418, 2012.
[15] Bilel Ben Romdhanne, Navid, Nikaein, Knopp Raymond, and BonnetChristian. OpenAirInterface large-scale wireless emulation platform andmethodology. In PM2HW2N 2011, 6th ACM International Workshopon Performance Monitoring, Measurement and Evaluation of Heteroge-neous Wireless and Wired Networks, October 31, 2011, Miami, Florida,USA, Miami, ÉTATS-UNIS, 10 2011.
[16] R. A. Berry and R. G. Gallager. Communication over fading channelswith delay constraints. IEEE Trans. Inf. Theor., 48(5):1135–1149, sep2006.
[17] I. Bettesh and S. Shamai. Optimal power and rate control for mini-mal average delay: The single-user case. Information Theory, IEEETransactions on, 52(9):4115 –4141, sept. 2006.
[18] Pierre Brémaud. Markov chains: Gibbs fields, Monte Carlo simulation,and queues. Texts in applied mathematics. Springer, 1999.
[19] G. Caire, G. Taricco, and E. Biglieri. Optimum power control overfading channels. Information Theory, IEEE Transactions on, 45(5):1468–1489, jul 1999.
[20] G. Caire and D. Tuninetti. The throughput of hybrid-ARQ protocols forthe Gaussian collision channel. Information Theory, IEEE Transactionson, 47(5):1971 –1988, july 2001.
[21] F. Capozzi, G. Piro, L.A. Grieco, G. Boggia, and P. Camarda. Down-link packet scheduling in lte cellular networks: Key design issues and asurvey. Communications Surveys Tutorials, IEEE, 15(2):678–700, 2013.
[22] C.C. Chai, Tjeng Thiang Tjhung, and Leng Chye Leck. Combined powerand rate adaptation for wireless cellular systems. Wireless Communi-cations, IEEE Transactions on, 4(1):6 – 13, jan. 2005.
[23] T.V.K. Chaitanya and E.G. Larsson. Outage-optimal power allocationfor hybrid arq with incremental redundancy. Wireless Communications,IEEE Transactions on, 10(7):2069 –2074, july 2011.
[24] V. Chandrasekhar, J. G. Andrews, and A. Gatherer. Femtocell net-works: A survey. IEEE Communications Magazine, 46(9):59–67, 2008.
Bibliography 131
[25] Yu Chen and Wei Wang. Machine-to-machine communication in LTE-A.In Vehicular Technology Conference Fall (VTC 2010-Fall), 2010 IEEE72nd, pages 1 –4, sept. 2010.
[26] H. Claussen. Performance of macro- and co-channel femtocells in ahierarchical cell structure, Sept. 2007.
[27] E. Dahlman, S. Parkvall, J. Sköld, and Per Beming. 3G Evolution HSPAand LTE for Mobile Broadband. Academic Press, 1st edition, 2007.
[28] E. Dahlman, S. Parkvall, and J. Sköld. 4G LTE/LTE-Advanced forMobile Broadband. Academic Press, 1st edition, 2011.
[29] H El Gamal, G Caire, and M. O. Damen. The mimo arq channel:Diversity-mutliplexing-delay tradeoff. IEEE Transactions on Informa-tion Theory, pages 3601–3621, August 2006.
[30] U. Erez, M.D. Trott, and Gregory W. Wornell. Rateless coding for gaus-sian channels. Information Theory, IEEE Transactions on, 58(2):530–547, 2012.
[31] EU FP7 Project LOLA (Achieving Low-Latency in Wireless Commu-nications). D4.3 Adaptive Modulation and Coding Scheme and HybridARQ Mechanism, v3.0, January 2013.
[32] Rizwan Ghaffar and Raymond Knopp. Interference suppression for nextgeneration wireless systems. In VTC 2009-Spring. IEEE 69th Vehicu-lar Technology Conference, 2009. April 26-29, 2009, Barcelona, Spain,April 2009.
[33] Amitabha Ghosh, Nitin Mangalvedhe, Rapeepat Ratasuk, BishwarupMondal, Mark Cudak, Eugene Visotsky, Timothy A. Thomas, Jef-frey G. Andrews, Ping Xia, Han-Shin Jo, Harpreet S. Dhillon, andThomas David Novlan. Heterogeneous cellular networks: From theoryto practice. IEEE Communications Magazine, 50(6):54–64, 2012.
[34] Arunabha Ghosh, Jun Zhang, Jeffrey G. Andrews, and Rias Muhamed.Fundamentals of LTE. Prentice Hall Press, 1st edition, 2010.
[35] A.J. Goldsmith. The capacity of downlink fading channels with variablerate and power. Vehicular Technology, IEEE Transactions on, 46(3):569–580, aug 1997.
[36] N. Gopalakrishnan and S. Gelfand. Rate selection algorithms for irhybrid arq. In Sarnoff Symposium, 2008 IEEE, pages 1 –6, april 2008.
[37] Aditya Gudipati and Sachin Katti. Strider: automatic rate adaptationand collision handling. SIGCOMM Comput. Commun. Rev., 41(4):158–169, aug 2011.
132 Bibliography
[38] H. Holma and A. Toskala. LTE for UMTS: Evolution to LTE-Advanced.John Wiley & Sons, 2011.
[39] Yichao Huang and B.D. Rao. An analytical framework for heterogeneouspartial feedback design in heterogeneous multicell ofdma networks. Sig-nal Processing, IEEE Transactions on, 61(3):753–769, 2013.
[40] Peter Iannucci, Jonathan Perry, Hari Balakrishnan, and Devavrat Shah.No symbol left behind: A link-layer protocol for rateless codes. In ACMMobiCom, Istanbul, Turkey, August 2012.
[41] IST-2003-507581 WINNER D1.3 version 1.0. Final usage scenarios.
[42] A. Khandekar, N. Bhushan, Ji Tingfang, and V. Vanghi. Lte-advanced:Heterogeneous networks. In Wireless Conference (EW), 2010 European,pages 978–982, 2010.
[43] Su Min Kim, Wan Choi, Tae Won Ban, and Dan Keun Sung. Optimalrate adaptation for hybrid arq in time-correlated rayleigh fading chan-nels. Wireless Communications, IEEE Transactions on, 10(3):968 –979,march 2011.
[44] Georgios P. Koudouridis and Hong Li. Distributed power on-off opti-misation for heterogeneous networks - a comparison of autonomous andcooperative optimisation. In CAMAD, pages 312–317. IEEE, 2012.
[45] Raymond Kwan, Cyril Leung, and Jie Zhang. Multiuser scheduling onthe downlink of an lte cellular system. Rec. Lett. Commun., 2008:3:1–3:4, jan 2008.
[46] A. Larmo, M. Lindström, M. Meyer, G. Pelletier, J. Torsner, andH. Wiemann. The LTE link-layer design. IEEE Communications Mag-azine, pages 52–59, April 2009.
[47] Lars Lindbom, Robert Love, Sandeep Krishnamurthy, Chunhai Yao,Nobuhiko Miki, and Vikram Chandrasekhar. Enhanced inter-cell in-terference coordination for heterogeneous networks in lte-advanced: Asurvey. CoRR, abs/1112.1344, 2011.
[48] A. Lioumpas, A. Alexiou, C. Anton-Haro, and P. Navaratnam. Ex-panding LTE for devices: Requirements, deployment phases and targetscenarios. Wireless Conference 2011 - Sustainable Wireless Technologies(European Wireless), 11th European, pages 1 –6, april 2011.
[49] M. Luby. Lt codes. In Foundations of Computer Science, 2002. Pro-ceedings. The 43rd Annual IEEE Symposium on, pages 271–280, 2002.
Bibliography 133
[50] Justin Manweiler, Sharad Agarwal, Ming Zhang, Romit Roy Choud-hury, and Paramvir Bahl. Switchboard: a matchmaking system formultiplayer mobile games. In Ashok K. Agrawala, Mark D. Corner, andDavid Wetherall, editors, MobiSys, pages 71–84. ACM, 2011.
[51] Ruben Merz and Jean-Yves Le Boudec. Performance evaluation of im-pulse radio uwb networks using common or private acquisition pream-bles. IEEE Transactions on Mobile Computing, 8:865–879, July 2009.
[52] Hyung G. Myung, Junsung Lim, and David J. Goodman. Single carrierfdma for uplink wireless transmission. Vehicular Technology Magazine,IEEE, 1(3):30–38, Sept. 2006.
[53] Navid Nikaein and Srdjan Krea. Latency for real-time machine-to-machine communication in LTE-based system architecture. WirelessConference 2011 - Sustainable Wireless Technologies (European Wire-less), 11th European, pages 1 –6, april 2011.
[54] Nokia Siemens Networks. Initial home nodeB coexistence simulationresults. 3GPP TSG-RAN WG4 R4-070902, June 2007.
[55] E. Ohlmer and G. Fettweis. Rate adaptation for interference cancelationreceivers in slowly time-variant mimo channels. In Sarnoff Symposium(SARNOFF), 2012 35th IEEE, pages 1 –5, may 2012.
[56] R. Palanki and Jonathan S. Yedidia. Rateless codes on noisy chan-nels. In Information Theory, 2004. ISIT 2004. Proceedings. Interna-tional Symposium on, pages 37–, 2004.
[57] S. Parkvall, A. FuruskaÌĹ andr, and E. Dahlman. Evolution of LTEtoward imt-advanced. Communications Magazine, IEEE, 49(2):84 –91,february 2011.
[58] Jonathan Perry, Peter A. Iannucci, Kermin E. Fleming, Hari Balakrish-nan, and Devavrat Shah. Spinal codes. In Proceedings of the ACM SIG-COMM 2012 conference on Applications, technologies, architectures,and protocols for computer communication, SIGCOMM ’12, pages 49–60, New York, NY, USA, 2012. ACM.
[59] J. G. Proakis. Digital Communications. McGraw–Hill, 4th edition, 2001.
[60] Changkyu Seol and Kyungwhoon Cheun. A statistical inter-cell in-terference model for downlink cellular OFDMA networks under log-normal shadowing and multipath Rayleigh fading. Communications,IEEE Transactions on, 57(10):3069 –3077, october 2009.
[61] Stefania Sesia, Issam Toufik, and Matthew Baker. LTE, The UMTSLong Term Evolution: From Theory to Practice. Wiley Publishing,2009.
134 Bibliography
[62] N. G. Shephard. From characteristic function to distribution function:A simple framework for the theory. Econometric Theory, (7):519–529,1991.
[63] A. Shokrollahi. Raptor codes. Information Theory, IEEE Transactionson, 52(6):2551–2567, 2006.
[64] Leszek Szczecinski, Ciro Correa, and Luciano Ahumada. Variable-rate retransmissions for incremental redundancy hybrid arq. CoRR,abs/1207.0229, 2012.
[65] T. Tabet and R. Knopp. Cross-layer based analysis of multi-hop wirelessnetworks. IEEE Transactions on Communications, 58(7), July 2010.
[66] A. Tajer, N. Prasad, and Xiaodong Wang. Fair rate adaptation in mul-tiuser interference channels. In Information Theory Proceedings (ISIT),2010 IEEE International Symposium on, pages 2083 –2087, june 2010.
[67] Daniela Tuninetti. Transmitter channel state information and repetitionprotocols in block fading channels. In Information Theory Workshop,2007. ITW ’07. IEEE, pages 505 –510, sept. 2007.
[68] T. Villa, R. Merz, and P. Vidales. Performance evaluation of OFDMAfemtocells link-layer in uncontrolled deployments. In Proceedings of Eu-ropean Wireless 2010, April 2010.
[69] Geng Wu, S. Talwar, K. Johnsson, N. Himayat, and K.D. Johnson.M2m: From mobile to embedded internet. Communications Magazine,IEEE, 49(4):36 –43, april 2011.
[70] P. Wu and N. Jindal. Performance of hybrid-arq in block-fading chan-nels: A fixed outage probability analysis. Communications, IEEETransactions on, 58(4), April 2010.
[71] K. Zheng, F. Hu, W. Xiangy, M. Dohler, and W. Wang. Radio resourceallocation in LTE-Advanced cellular networks with M2M communica-tions. IEEE Communications Magazine, accepted for publication, 2012.
